Gradient-index, or GRIN, lenses have long fascinated engineers because they bend and focus waves not through curved surfaces, as conventional optics does, but through spatially varying dielectric properties distributed throughout the volume of the material. Instead of a shaped piece of glass, a GRIN lens is essentially a structured medium whose effective refractive index changes gradually from point to point. When a wave propagates through such a medium, its wavefront is progressively reshaped, allowing the lens to collimate, focus, or steer beams with remarkable freedom. For compact beam-steering systems, particularly those needed in satellite communications, radar, and next-generation wireless networks, this kind of wavefront control in a flat or modestly sized package is extremely attractive.

The difficulty has always been the inverse problem. Forward design is straightforward: given a specified permittivity distribution, electromagnetic solvers can predict exactly how waves will behave. Inverse design asks the opposite question. Given a desired field transformation, what volumetric permittivity distribution produces it? For most practical devices, this question has been answered through voxel-by-voxel topology optimization, in which the design volume is divided into thousands or millions of tiny cells, each of which is iteratively adjusted. Every iteration demands a full-wave electromagnetic simulation, and a converged design can require thousands of such simulations. On ordinary computing hardware, this process can stretch from days into weeks, and the resulting designs are often highly resonant, meaning they perform well only in a narrow frequency band and are exquisitely sensitive to fabrication tolerances.

The Siena and Inkbit team took a fundamentally different route. Rather than optimizing an arbitrary voxel map, they restricted the search space to smooth GRIN profiles appropriate for propagation-based devices. This constraint is not a limitation but a strategic advantage. Smooth, slowly varying index profiles are exactly the regime in which geometrical optics provides an accurate and computationally cheap description of wave behavior. By optimizing the geometrical optics propagation equations directly, the researchers sidestepped the need for repeated full-wave simulations entirely. The optimization problem collapses from a monstrous high-dimensional search into a tractable one that runs in minutes on a standard computer. The physics of propagation, rather than raw computational power, does the heavy lifting.

The mathematical logic behind the approach is elegant. Geometrical optics treats waves as rays that follow paths determined by the local refractive index, and the accumulated phase along each ray dictates how the wavefront evolves. By parameterizing the permittivity distribution with a modest number of smooth basis functions, the designers created a differentiable mapping from the lens parameters to the resulting field transformation. Gradient-based optimization then adjusts those parameters until the predicted output field matches the target, whether that target is a focused spot, a collimated beam, or, in the demonstrated case, a set of independently steered beams corresponding to different feed positions. Because the optimization operates on smooth profiles, the resulting designs are inherently broadband: a gradual index variation does not suddenly stop working when the frequency shifts, unlike resonant structures tuned to a single operating point.

To prove that the method delivers real hardware rather than pretty simulations, the team fabricated a multi-feed GRIN lens using additive manufacturing. The internal architecture is based on a three-dimensional gyroid, one of the most celebrated geometries in modern materials science. A gyroid is a triply periodic minimal surface, a structure that divides space into two intertwined, self-connected labyrinths with no straight edges or flat faces. By controlling the local density and dimensions of the gyroid lattice, the effective dielectric constant of the material can be tuned continuously across the lens volume, translating the smooth permittivity profile computed by the optimizer into a physical object. Additive manufacturing is the enabling technology here, since no conventional machining process could produce such an intricate, spatially graded internal scaffold.

The experimental results are striking. The fabricated lens supports multiple feed ports, each of which launches a beam in a different direction, effectively turning a single passive structure into a multi-beam former. As the feed is switched, the beam scans continuously across a ±35-degree range, and the measured performance holds across a 33 percent fractional bandwidth, a figure that comfortably encompasses many practical communication and sensing bands. The researchers report low scan loss, meaning the beam does not weaken appreciably as it moves away from boresight, and excellent beam fidelity, meaning the radiated pattern retains its shape and sidelobe behavior across the scanning range. These are precisely the attributes that matter in real phased-array and multibeam antenna systems, where scan loss and pattern distortion have historically limited performance.

The significance of the demonstration extends beyond the specific device. Multi-beam formers are workhorses of satellite ground stations, 5G and future 6G base stations, and high-throughput communication links, where a single aperture must serve many users or satellites simultaneously. Conventional solutions rely on phased arrays with hundreds of individually controlled elements, each requiring amplifiers, phase shifters, and calibration, or on quasi-optical systems such as Rotman lenses and reflectarrays that involve their own design compromises. A passively operated GRIN lens that steers beams simply by selecting among fixed feeds offers a dramatically simpler architecture with no active electronics in the beam-forming path, reducing cost, power consumption, and failure modes.

Equally important is what the method implies for the design workflow of metamaterial devices in general. The gap between an electromagnetic specification and a printable file has been the central obstacle to scalable wavefront engineering. Topology optimization, for all its power, produces designs whose complexity is entangled with the simulation cost of producing them, and whose narrowband resonant character often clashes with real-world bandwidth requirements. By anchoring the inverse problem in geometrical optics and smooth profiles, the new framework establishes what the authors describe as a direct pathway from electromagnetic functionality to manufacturable metamaterial devices. The optimization is fast enough to be iterative in a design office, and the output is, by construction, compatible with additive manufacturing processes that can realize graded effective media.

The work was carried out within the framework of the Huawei–University of Siena Joint Laboratory, a collaboration that reflects the growing industrial appetite for beam-steering hardware as wireless systems push toward higher frequencies and denser networks. The article, authored by Ilir Gashi, Scott Twiddy, Zachary Nelson, Stefano Maci, and Matteo Albani, with Twiddy and Nelson contributing equally, was published open access on 26 September 2026. As additive manufacturing continues to mature, allowing ever finer control over graded internal architectures, the combination of physics-informed inverse design and freeform fabrication points toward a future in which antennas and lenses are no longer designed component by component but computed as a whole, with the desired electromagnetic behavior specified first and the material emerging as the answer. For a field that has spent decades choosing between the speed of approximate models and the accuracy of exhaustive simulation, a method that achieves wideband, experimentally verified performance in minutes of computation marks a genuine turning point.

Subject of Research: Inverse design of gradient-index metamaterials for wideband multi-beam forming

Article Title: Inverse design of metamaterials for wideband multi-beam former

Article References: Gashi, I., Twiddy, S., Nelson, Z., Maci, S., & Albani, M. (2026). Inverse design of metamaterials for wideband multi-beam former. Communications Engineering. https://doi.org/10.1038/s44172-026-00787-1

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00787-1

Keywords: metamaterials, inverse design, gradient-index lens, beam steering, additive manufacturing, gyroid, multi-beam former, electromagnetics, geometrical optics, antenna engineering, topology optimization, wavefront engineering

Carbon nanotubes are hollow cylinders of carbon atoms with walls only one atom thick in their most extreme form, yet they possess remarkable mechanical strength, thermal conductivity and electrical properties. When millions of them grow simultaneously on a substrate, they rise in parallel like blades of grass, forming what researchers call an array or, more evocatively, a nanotube forest. The way this forest stands, how tall it grows, how tightly packed the individual tubes are, and how well they align with one another, determines how the material performs in real devices. Thermal interface materials rely on dense, well-aligned arrays to channel heat away from hotspots in electronics. Energy storage electrodes benefit from high surface area and controlled porosity. Flexible electronics demand forests that can bend and recover without collapsing. In each case, the macroscopic usefulness of the material traces directly back to those three structural descriptors: height, density and alignment.

The trouble is that these descriptors cannot be tuned independently. Growing nanotube forests typically relies on chemical vapour deposition, a process in which a carbon feedstock gas decomposes on catalytic nanoparticles, allowing tubes to grow upward from the surface. Water-assisted chemical vapour deposition, the variant used in this study, introduces controlled amounts of water vapour to keep the catalyst active and extend growth. But the process is a web of interdependent variables. Raising the growth temperature might accelerate catalysis and increase height, while simultaneously altering how quickly the catalyst particles sinter or deactivate, which in turn changes tube density. Adjusting gas flows, treatment times, or the timing of water injection can shift several structural features at once. A modification that nudges height toward its target may push density or alignment further away, forcing researchers into laborious trial-and-error cycles. The scale of the search space is staggering: with eight processing variables and even ten candidate settings per variable, the number of possible combinations reaches one hundred million.

To tame this complexity, the research team, drawn from Huazhong University of Science and Technology and Beihang University, developed an AI-assisted inverse-design framework. Inverse design flips the usual logic of materials synthesis. Instead of running an experiment and measuring what comes out, the researcher specifies the outcome first, the desired height, density and alignment, and asks an algorithm to find the processing conditions most likely to deliver it. The team began by assembling an experimental database of two hundred distinct recipes for water-assisted chemical vapour deposition, each recording the full set of processing conditions alongside the structural features of the resulting array. A neural network then absorbed this dataset, learning the intricate mapping from processing space to structural space, including the nonlinear couplings that make manual optimization so difficult.

On top of this learned model, the researchers deployed an optimization algorithm that searches processing space in reverse, proposing recipes predicted to hit a prescribed structural target. Crucially, the framework was designed not to find one answer but many. For each of three distinct structural targets, repeated computational searches generated two hundred candidate solutions. Rather than accepting the single best-scoring recipe, the team deliberately selected three candidates whose processing conditions differed substantially from one another. This is where the mountain analogy becomes more than a metaphor: the optimization landscape contains many peaks that satisfy the same structural criteria, and a navigator equipped with a learned map can explore routes far from the well-worn trail surrounding a familiar recipe.

The experimental validation was rigorous. Each of the three selected recipes for each target was tested in three independent growth batches, producing twenty-seven validation runs in total. Of these, twenty-six achieved structural descriptor matching scores above ninety-five percent, meaning the measured height, density and alignment of the grown arrays came within a few percent of the prescribed targets. The consistency across independent batches matters as much as the scores themselves, because it demonstrates that the AI-identified recipes are not statistical flukes but reproducible synthesis protocols. For all three targets, the framework had successfully converted a specification into working laboratory instructions, and it had done so through multiple, demonstrably different routes.

Yet the story acquires its most intriguing twist at smaller length scales. When the researchers examined the nanotubes themselves under electron microscopy, the forests that looked identical at the array level turned out to be built from different trees. For one target, the three successful routes produced arrays with closely matched overall features, but the mean outer diameters of the individual nanotubes measured approximately 3.8, 4.8 and 7.2 nanometres respectively. The wall structures of the tubes, whether single-walled or multi-walled and how many concentric layers they contained, also differed between routes. Similar forests, in other words, need not contain identical trees. This observation carries practical weight: applications sensitive to nanotube diameter or wall number, such as electronic transport or optical absorption, may respond differently to arrays that satisfy the same coarse structural specification.

The broader implication is a shift in what inverse synthesis aims to accomplish. Traditional optimization seeks a single best recipe, but this work demonstrates that the goal can be reframed as identifying a family of experimentally accessible alternatives, each satisfying the same structural target while differing in the practical demands they place on a laboratory. One route might use a temperature that a particular furnace reaches easily; another might favour gas flows that are cheaper or safer. The AI framework functions as a navigation aid built from accumulated experimental knowledge, and its learned relationship between processing and structure can be reused to screen additional targets without repeating the full experimental investment. The initial database of two hundred recipes remains a significant cost, and extending the approach to other catalysts, feedstocks or growth methods would require new experimental data and model retraining, but within its trained domain the framework offers a reusable map of synthesis space.

For a field that has long treated nanotube growth as something of an artisanal craft, the study offers a glimpse of a more flexible future, one in which engineers specify the structure they need and choose among several validated paths to reach it, constrained only by the equipment and materials at hand. The work, led by first author Lei Zhu, a doctoral student at Huazhong University of Science and Technology, with Professor Ming Xu as corresponding author, suggests that the one hundred million possible combinations of processing variables are not an obstacle but an opportunity: hidden within that vast space lie many recipes for the same material, and artificial intelligence is now capable of finding them, testing them against reality, and revealing that even at the nanoscale, there is more than one way to grow a forest.

Subject of Research: AI-assisted inverse design of carbon nanotube array synthesis by chemical vapour deposition

Article Title: AI finds different recipes for carbon nanotube forests with matching structures

Article References: AI finds different recipes for carbon nanotube forests with matching structures. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: carbon nanotubes, inverse design, artificial intelligence, chemical vapour deposition, nanomaterials, machine learning, materials synthesis, nanotube arrays, neural networks, National Science Review, nanotechnology, structural optimization

Now a team at China University of Petroleum (Beijing), working with Shandong Kegu Jiequan Technology Co., Ltd., has built an artificial intelligence framework that learns to read fuel molecules the way a chemist would, capturing the structural subtleties that determine octane behaviour. Writing in the journal ENG. Chem. Eng., the researchers describe an interpretable multimodal molecular representation model that predicts RON for both pure compounds and complex blends with high accuracy, and then turns the prediction problem on its head to enable inverse fuel design, the task of finding blend compositions that hit a target octane number before a single drop is mixed in the laboratory.

The core innovation lies in how the model represents each molecule. Rather than relying on a single descriptor, the framework integrates three complementary views of molecular structure. The first is a graph neural network embedding, in which atoms become nodes and chemical bonds become edges, allowing the network to learn the topological relationships that define a molecule’s skeleton. The second is a set of MACCS fingerprints, binary strings that encode the presence or absence of predefined substructure fragments, giving the model a chemist’s vocabulary of functional groups and ring systems. The third consists of molecular descriptors, numerical summaries of physicochemical properties, selected through a residual-guided strategy that identifies which descriptors add explanatory power beyond what the learned representations already capture.

For pure-component RON prediction, the trimodal model achieved a coefficient of determination, R², of 0.9373 and a mean absolute error of 4.04 octane units on the test set. Ablation experiments, in which individual information channels were systematically removed, revealed that the graph topology contributed the strongest signal, while the MACCS fingerprints and descriptors supplied complementary information that sharpened the predictions. The result is a model that does not merely memorise correlations but assembles a genuinely multi-perspective picture of what makes a molecule knock-resistant.

What sets the framework apart from many black-box models is its interpretability. The graph neural network employs an atom-level attention mechanism that visualises which local structural environments the model emphasises when making a prediction. The patterns it highlights are strikingly consistent with decades of empirical knowledge about structure-octane relationships. In alkanes, the model concentrated on branching sites, the structural features long known to boost octane quality. In cycloalkanes, it attended to substitution sites; in olefins, to the reactive double-bond regions; and in aromatics, to the connection points between side chains and the aromatic ring. For fuel chemists, this alignment between machine attention and chemical intuition is a crucial trust signal, indicating that the model has internalised genuine structure-property physics rather than exploiting dataset artefacts.

The real challenge, however, lies in mixtures. A fuel blend is not simply a collection of independent molecules; components interact in ways that shift the effective octane rating away from any weighted average. The researchers tackled this by transferring the trained pure-component encoder to the mixture problem. For a mixture containing multiple components, the embedding vector of each component was weighted according to its mole fraction and combined in the latent space, the abstract high-dimensional space where the neural network represents molecular structure. This composition-weighted latent representation was then fed into an XGBoost regressor, a gradient-boosted tree model, to produce the final RON prediction.

The performance gain over conventional methods was substantial. The first-order latent-space mixing model achieved an R² of 0.9736 and a mean absolute error of 1.46 on the mixture test set, dramatically outperforming the linear blending baseline, which managed only an R² of 0.7501 with a mean absolute error of 4.83. The comparison makes the failure of linear blending rules vivid: the AI approach cut prediction error by roughly seventy percent. Interestingly, when the team added second-order interaction terms designed to capture pairwise component interactions, the improvement was not significant. This suggests that the first-order model already captured the primary composition-dependent variation within the learned latent space, implying that the neural embeddings themselves encode much of the interaction chemistry that linear rules miss.

Prediction is only half the story. The researchers then demonstrated that the model could support fuel formulation design, the inverse problem of specifying a blend that meets a target octane constraint. Using a stochastic sampling search method, they identified feasible ternary blending compositions satisfying target RON requirements across four case studies. In every case, known formulations reported in the literature fell within the predicted feasible solution space, confirming that the computational search does not exclude chemically realistic answers and can genuinely guide formulation work. For refiners and fuel developers, this means candidate blends can be screened computationally, reserving laboratory time and materials for the most promising candidates rather than an exhaustive trial-and-error campaign.

The broader significance of the work is methodological. It demonstrates that molecular representations learned from pure components can be effectively transferred to mixture property prediction, a strategy that could spare researchers from assembling large, costly mixture datasets. It also establishes latent-space composition weighting as a promising general approach for mixture property modelling, one that respects the nonlinear reality of blending behaviour without requiring explicit knowledge of every possible interaction. Because the underlying encoder is trained on pure compounds, data that are far more abundant and cheaper to obtain, the framework lowers the barrier to accurate mixture modelling across the fuel industry.

The authors view the current model as a foundation rather than a finished product. The natural next step is multi-objective optimisation, in which octane number is balanced against additional fuel properties such as vapour pressure, density and viscosity, all of which constrain what a practical fuel formulation can look like. As transportation fuels evolve toward novel blends, oxygenated components and synthetic hydrocarbons, tools that can predict and design fuel properties computationally are likely to become indispensable. This study, published with the DOI 10.1007/s11705-026-2702-7, offers a concrete demonstration that interpretable machine learning can move fuel science from measurement toward design, turning the octane number from a laboratory bottleneck into a variable that engineers can dial in.

Subject of Research: Multimodal machine learning for research octane number prediction and inverse fuel blend design

Article Title: Multimodal AI model predicts octane number of fuel blends with high accuracy, enables inverse fuel design

Article References: Multimodal AI model predicts octane number of fuel blends with high accuracy, enables inverse fuel design. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: research octane number, multimodal AI, graph neural networks, fuel blending, inverse design, XGBoost, molecular descriptors, MACCS fingerprints, latent space mixing, fuel formulation, machine learning, chemical engineering

The core problem the team set out to solve is one that anyone who has fabricated a microchip will recognize. Photonic neural networks promise enormous gains in speed and energy efficiency because light can perform massive numbers of multiply-and-accumulate operations in parallel, at the speed of propagation and with almost no heat dissipation. But before such a chip can do useful work, its tunable elements — phase shifters, couplers, and other adjustable components — must be set to precise values. Traditionally, those values are found by training a digital twin of the circuit in software. That twin is built from physical models of the device, and no model is perfect. Tiny discrepancies between the simulated chip and the fabricated one, introduced by manufacturing tolerances, thermal drift, and unmodeled nonlinearities, accumulate through the layers of a network and can destroy its accuracy. Re-modeling every chip individually is computationally expensive and often impractical at scale.

INSPIRE sidesteps the digital twin entirely. The method works by measuring, on the chip itself, the full complex optical fields of the bidirectional modes that propagate through the circuit — that is, both the amplitude and the phase of the light traveling forward and backward through the device. The key enabling technology is what the authors call on-chip synthetic time-reversal holography. Holography is a well-established technique for recording complete wavefronts, including phase information that ordinary intensity detectors miss. By synthesizing a time-reversed counterpart of the light field on the chip, the system can effectively send light backward through the circuit and interfere it with reference beams, extracting the field information needed to compute gradients. Those gradients — the same mathematical objects that drive learning in conventional deep networks — are then used to update the chip’s tunable parameters directly, so the physical system and the learning algorithm are one and the same process.

What makes the approach especially powerful is its generality. Earlier demonstrations of in situ training in photonics were often tailored to a specific architecture, such as a particular mesh of interferometers or a diffractive stack. INSPIRE, by contrast, is topology-agnostic: it does not care how the optical circuit is wired together. As long as the device has tunable elements and the bidirectional fields can be measured, the method applies. That compatibility with diverse optical circuits — from integrated interferometer meshes to metasurface-based processors — is what the authors mean by calling it a generalized training framework, and it is what separates this work from the growing but fragmented family of physical training techniques that have appeared in recent years.

The experimental results are striking. In benchmark demonstrations, the team trained matrices directly on the photonic hardware and achieved a relative error of just 0.26 percent against the target values — a level of precision that indicates the physical gradient computation is not merely a rough approximation but a faithful replacement for software-based optimization. Perhaps more surprising is what the researchers achieved when they pushed the method through a scattering medium, a chaotic optical environment that scrambles light in ways that are notoriously difficult to model. Using INSPIRE, they trained matrices larger than the number of native tunable elements on the chip, effectively extracting more computational capacity from the hardware than its component count would suggest possible. Because the training happens in situ, the scattering and imperfections of the medium become part of the computation rather than obstacles to it.

The team then extended the framework into the realm of meta-learning — the science of teaching systems how to learn quickly. Meta-photonic circuits trained with INSPIRE were able to perform in situ meta-learning, adapting to new tasks with remarkable efficiency. The reported numbers are eye-catching: a 251-fold compression of the model and a 136-fold acceleration in task-specific training compared with conventional approaches. In practical terms, this means a single photonic circuit can be prepared so that it snaps to a new task after minimal additional training, a property known as single-shot photonic learning. For applications where a device must be reconfigured on the fly — a sensor that changes environment, a camera that switches between imaging modes — that kind of rapid adaptability could be transformative.

The broader context explains why this demonstration has generated excitement well beyond the photonics community. Deep learning’s explosive growth has collided with the slowing of Moore’s law, and the energy cost of training and running large models on digital hardware has become a first-order concern for the industry. Optical computing has long been proposed as an escape route, and laboratory demonstrations have shown photonic processors performing inference with extraordinary efficiency — some operating with less than one photon per multiplication. But the training bottleneck has persisted: if every chip must be modeled and calibrated in software, the promise of cheap, mass-producible optical AI hardware remains out of reach. A training method that lives on the chip itself, that is indifferent to the circuit’s design, and that absorbs fabrication errors into the learning process, addresses that bottleneck head-on.

INSPIRE also fits into a rapidly evolving landscape of physical training research. Recent years have seen in situ backpropagation demonstrated in photonic meshes, forward-only training schemes that avoid sending signals backward through hardware, and theoretical frameworks for training physical neural networks of many kinds. Each of these advances has chipped away at the problem, but most have carried architectural constraints or required specialized hardware modifications. The holographic field-measurement approach at the heart of INSPIRE is notable because it treats the optical circuit as a black box whose bidirectional response can be interrogated, making the training procedure a property of the measurement scheme rather than of the device design. The authors have also released the software code for the model and training procedure through Zenodo, which should make it easier for other groups to adopt and extend the method.

There are, of course, questions that future work must answer. The demonstrations reported here, while impressive, were carried out on laboratory-scale systems, and scaling to the very large networks that commercial applications would demand will require careful engineering of the holographic measurement apparatus and the on-chip tunable elements. The speed at which gradients can be measured and applied, the stability of the trained parameters over time and temperature, and the integration of the training hardware alongside the computing hardware on a single chip are all open engineering challenges. The energy accounting of the full training pipeline, including the lasers and detectors involved in the holographic measurements, will also need to be quantified as the technology matures.

Even so, the trajectory is clear. A photonic chip that can measure its own internal light fields, compute its own gradients, and update its own parameters represents a meaningful step toward adaptive, self-configuring optical processors. If the approach scales, the consequences could reach far beyond data centers: smart sensors that learn their environment at the point of capture, imaging systems that retrain themselves in microseconds, and AI hardware whose energy footprint is measured in milliwatts rather than megawatts. The authors describe their work as a practical route toward adaptive and efficient intelligent photonic systems, and with error rates below one percent and training that survives scattering media, that route now looks considerably more navigable than it did before.

Subject of Research: In situ physical gradient descent training of photonic neuromorphic integrated circuits

Article Title: Photonic neuromorphic learning via generalized in situ physical gradient descent

Article References: Zhou, T., Zhao, Y., Li, S., Shao, G., Huang, R., & Fang, L. (2026). Photonic neuromorphic learning via generalized in situ physical gradient descent. Nature Computational Science. https://doi.org/10.1038/s43588-026-01057-y

Image Credits: AI Generated

DOI: 10.1038/s43588-026-01057-y

Keywords: photonics, neuromorphic computing, in situ training, gradient descent, holography, optical neural networks, meta-learning, photonic integrated circuits, energy-efficient computing, scattering media, inverse design, Nature Computational Science

The appeal of high-entropy alloys as catalysts stems from several intertwined effects. Their high configurational entropy stabilizes single-phase structures that would otherwise separate into distinct compounds, while the so-called cocktail effect produces synergistic interactions among elements that no individual component displays on its own. Their electronic structures can be tuned continuously by adjusting composition, and their lattice distortion and sluggish diffusion often confer exceptional structural stability under harsh reaction conditions. Experiments have demonstrated their promise across ammonia decomposition, hydrogen evolution, oxygen reduction, carbon dioxide reduction, and nitrate-to-ammonia conversion, among other reactions central to a sustainable energy economy.

Yet the very feature that makes these materials exciting also makes them nearly impossible to explore by brute force. With dozens of candidate elements and essentially continuous control over mixing ratios, the compositional space of a five-element alloy alone spans numbers of candidates that dwarf any conceivable experimental or even computational screening campaign. Traditional density functional theory calculations, which treat each composition and surface configuration individually, are far too slow to map such a landscape. The review’s authors argue that this is precisely where machine learning has become indispensable: predictive models trained on a manageable set of calculations or experiments can interpolate across the vast remaining space, turning an intractable search into a guided exploration.

The methodological toolkit described in the review spans a wide spectrum of sophistication. At the simpler end sit classical regression techniques, including linear regression, kernel ridge regression, and gradient-boosted decision trees, which remain workhorses for predicting catalytic descriptors such as adsorption energies from composition-based features. Neural networks, from early dimensionality-reduction architectures to modern deep learning models, capture more complex nonlinear relationships. Graph neural networks and attention-based architectures have proven particularly powerful because they encode the local atomic environment directly, allowing models to distinguish between different adsorption sites on an alloy surface that would look identical to composition-only descriptors. Interpretable machine learning approaches have further helped researchers extract physical insight, such as electronic descriptors tied to local chemical environments, rather than treating models as black boxes.

A particularly important strand of work concerns the d-band center, the electronic-structure descriptor that underpins the classic Sabatier principle linking adsorption strength to catalytic activity. Recent studies have produced general models capable of predicting d-center positions across multi-principal-element alloys, and researchers have uncovered unusual Sabatier behavior on high-entropy surfaces where the conventional volcano-shaped activity relationship breaks down or shifts. Neural network approaches have also been used to decouple ligand effects, which arise from electronic interactions among neighboring elements, from coordination effects tied to the local atomic geometry, a distinction that is crucial for rational design but nearly impossible to isolate experimentally.

Machine learning interatomic potentials represent another transformative advance. These models learn the potential energy surface directly from quantum-mechanical calculations and then run atomistic simulations at a tiny fraction of the computational cost, achieving near-density-functional-theory accuracy at scales approaching classical force fields. Equivariant message-passing architectures of the kind embodied in modern frameworks have pushed accuracy to new levels, and benchmarking efforts are now establishing how well these potentials handle the chemical complexity of multicomponent alloys. Combined with Monte Carlo sampling, including distributed implementations that scale to trillions of atoms on AI accelerators, these tools allow researchers to simulate surface segregation, chemical short-range order, and nanostructure formation in high-entropy particles, phenomena that govern which atoms actually sit at the catalytically active surface.

The review highlights concrete applications where machine learning has already accelerated discovery. Multi-objective optimization over vast composition spaces has identified high-entropy electrocatalysts for carbon dioxide and carbon monoxide reduction that would have been difficult to find by intuition. Machine-learning-guided screening has helped stabilize ruthenium in multicomponent alloys for acidic oxygen evolution, a notoriously corrosive environment where catalyst durability is a bottleneck. High-throughput experimentation paired with data-driven strategies has delivered efficient hydrogen evolution catalysts, while interpretable deep graph attention learning has been used to design high-entropy electrocatalysts with optimized adsorption properties. Generative artificial intelligence has even enabled inverse design, in which researchers specify a target property and the model proposes compositions likely to achieve it, an approach demonstrated for high-entropy catalyst discovery in a recent Nature Synthesis study.

Perhaps the most forward-looking section of the review concerns large language models. These systems, originally developed for text, are being repurposed to mine the scientific literature at scale, extracting composition-property relationships from millions of published papers, a strategy previously shown to enable the design of ultrahigh-entropy alloys. In catalysis, language models are now being used for knowledge extraction, hypothesis generation and validation, and as orchestrating agents that tie together databases, quantum calculations, machine learning surrogates, and robotic laboratories into integrated design workflows. Retrieval-augmented generation, which grounds model outputs in retrieved source documents, has been applied to propose high-entropy catalyst candidates, and multimodal models that combine textual and graphical understanding of adsorption configurations are extending these capabilities further.

The authors are candid about the challenges that remain. Data scarcity and quality are chronic problems: experimental datasets for high-entropy catalysts are small, heterogeneous, and often reported without the standardized metadata needed for machine learning. Models trained on one family of alloys frequently fail to transfer to another, and the distribution of atomic environments in these materials is so broad that extrapolation beyond the training domain remains risky. Surface segregation means the composition of the working catalyst may differ substantially from the bulk, complicating any composition-based prediction. Interpretable models trade accuracy for transparency, while accurate deep models can obscure the physics. Benchmarking standards, community datasets such as the Open Catalyst collections, and careful validation against experiment are emerging as essential safeguards.

The trajectory outlined in the review points toward closed-loop, autonomous discovery: language-model agents that read the literature and propose hypotheses, machine learning potentials that simulate candidate surfaces atom by atom, generative models that invert property targets into compositions, and self-driving laboratories that synthesize and test them in rapid iteration. If that integration matures, the staggering compositional space of high-entropy alloys, once a barrier, becomes the field’s greatest asset, an almost limitless reservoir of catalytic solutions for hydrogen production, carbon conversion, and green chemistry. The review’s message is that the tools to explore that reservoir now exist; the task ahead is to make them reliable, interpretable, and accessible enough that rational design becomes the norm rather than the exception.

Subject of Research: Machine learning methods for designing high-entropy alloy catalysts

Article Title: Machine learning for high-entropy catalysts: methods and applications

Article References: Chen, H., Pei, Z., & Liu, X. (2026). Machine learning for high-entropy catalysts: methods and applications. Journal of Materials Science. https://doi.org/10.1007/s10853-026-13834-1

Image Credits: AI Generated

DOI: 10.1007/s10853-026-13834-1

Keywords: high-entropy alloys, machine learning, catalysis, electrocatalysis, density functional theory, machine learning interatomic potentials, graph neural networks, large language models, adsorption energy, surface segregation, hydrogen evolution reaction, inverse design

The logic of singlet fission is deceptively simple. For the process to be energetically viable, the energy of the first excited singlet state, S1, must be at least equal to twice the energy of the lowest triplet state, T1, so that one singlet exciton can divide into two triplets without losing energy. Ideally, T1 should also sit above the bandgaps of established semiconductors such as silicon, GaAs or CdTe — roughly 1.1 to 1.5 electronvolts — so that the triplet energy can be injected efficiently into a device. Tetracene, with a triplet energy near 1.25 eV, and pentacene, near 0.85 eV, meet these requirements naturally. Anthracene, benzene and naphthalene do not: their triplet levels are too high, making fission endothermic for the bare hydrocarbon skeletons, and no examples of singlet fission in benzenes or naphthalenes have been reported at all.

The Granada group reasoned that strategic substitution could bend the energy levels of even the smallest acenes into the right configuration. The obstacle is scale. Even restricting the search to six functional groups placed at six positions across the family generates around 10^8 candidate structures; lifting those restrictions and allowing hundreds of substituents at up to six positions pushes the count toward 10^19. No laboratory, and no brute-force quantum chemical calculation, could enumerate such a space. The answer, the researchers argue, is inverse design: instead of synthesizing a molecule and measuring its properties, one specifies the desired property and works backward to the structures most likely to deliver it.

Their pipeline rests on two pillars. The first is a predictive model: a gated recurrent unit network, a type of recurrent neural network implemented in PyTorch, that reads a molecule as a sequence and converts its connectivity into a fixed-size mathematical representation. Crucially, the electronic character of each substituent is encoded not by an opaque identifier but by its classical Hammett sigma constant, the century-old parameter of physical organic chemistry that quantifies whether a group withdraws or donates electron density. This choice gives the model chemical transparency and, because many different groups share similar sigma values, allows it to generalize to substituents it has never seen by interpolating within the continuous descriptor space.

The second pillar is data. The team computed the S1 and T1 energies of nearly 30,000 substituted acenes — benzenes, naphthalenes, anthracenes, tetracenes and pentacenes — using time-dependent density functional theory at the B3LYP/6-31G(d) level, a method previously shown to reproduce acene excitation energies to within about 0.1 to 0.2 eV of experiment. The training set used six substituents spanning sigma values from −0.66 to 0.78, including fluorine, amino, hydroxyl, methyl, cyano and nitro groups, with up to four substituents per molecule. The network, a two-layer unidirectional GRU with a hidden state of dimension 32, explicitly averages its predictions over all symmetry-equivalent orderings of each molecule — twelve permutations for benzene’s dihedral group, four for linear acenes — so that the result is invariant to how the structure is written down. Typical prediction errors landed between 1 and 3 percent for S1 and 1 to 6 percent for T1, and the model remained reliable even for penta- and hexa-substituted compounds outside its training regime.

On top of the predictor, the researchers mounted two optimization engines. Genetic algorithms evolve populations of candidate molecules, treating the network’s predicted energies as a fitness function, while particle swarm optimization sends a swarm of solutions exploring the sigma-constant landscape. Both can be constrained — by core size, substituent type, count and symmetry — so that the suggestions remain compatible with real synthetic chemistry. Every candidate proposed by the algorithms was then validated with full DFT calculations. The team emphasizes that perfect quantitative accuracy is not the point; what matters is whether the search returns molecules that genuinely satisfy the fission criterion, and it does.

The results split into two categories. For tetracenes and pentacenes, the algorithms rediscovered many known singlet fission chromophores but also revealed that the absorption edge, set by S1, can be tuned at will rather than being pinned to the values of the unsubstituted parents — an attractive degree of freedom for matching solar spectra. Far more striking is what happened with the smaller cores. The search produced validated benzene, naphthalene and anthracene derivatives meeting the S1 ≥ 2T1 condition, several with triplet energies above 1.4 eV, the threshold proposed for efficient charge injection into silicon. These are structures that chemical intuition alone would almost certainly never have surfaced, and the authors describe them as imposing a new paradigm for rational candidate identification.

The particle swarm stage pushed the exploration further. Treating sigma constants as a continuum rather than a fixed menu, the swarm could propose new electronic profiles and then map them onto real functional groups drawn from a list of roughly 530 known Hammett constants, expanding the accessible space to the full 10^19 structures. Starting from a single seed molecule, the algorithm mutated one substituent at a time or several simultaneously, sometimes replacing every sigma constant in the molecule while leaving the target property essentially unchanged — a demonstration that the optimizer perceives subtle electronic compensations invisible to a human designer. Even seeds that initially failed the energetic criterion were steered toward viable solutions. Conformational searches with the CREST software confirmed that, for the highlighted candidates, the dominant conformer reproduced the ensemble-averaged energy levels, making the predictions robust to molecular flexibility.

Perhaps most valuable is that the black box can be opened. Analyzing the anthracene candidates, the team identified recurring substitution patterns: a strong preference for modifying the internal ring positions, either through donor–acceptor pairings or through moderate electron-withdrawing groups flanked by donors positioned to enable hydrogen bonding. Armed with these extracted rules, the researchers designed new anthracene derivatives by hand and confirmed by TD-DFT that the targeted motifs reliably deliver molecules meeting the fission criterion — including structures with no precedent in the literature. The model, trained on just six substituents, had learned transferable design principles.

The protocol is freely accessible as an interactive tool at alba.ugr.es/acene/, and the authors suggest it can be extended to other molecular scaffolds and photophysical targets. Singlet fission remains a complex solid-state phenomenon, and the energetic criterion is necessary but not sufficient: crystal packing, intermolecular coupling and higher excited states all matter, so the AI screen is best understood as a powerful first-pass filter that hands experimentalists a shortlist of promising regions of chemical space. Even so, the demonstration that machine learning coupled to classical physical organic descriptors can conjure singlet fission candidates from benzene — the most iconic molecule in chemistry, discovered two centuries ago — shows how inverse design is beginning to turn the astronomically large universe of possible molecules into a searchable map.

Subject of Research: AI-driven inverse design of singlet fission chromophores in the acene family

Article Title: Artificial Intelligence‐Driven Inverse Design of Singlet Fission Candidates in the Acene family

Article References: Uceda, R. G., Pérez‐Cañedo, B., Míguez‐Lago, S., Cruz, C. M., Torres, J. J., Núñez, O., Álvarez de Cienfuegos, L., Mota, A. J., Miguel, D., & Cuerva, J. M. (2026). Artificial Intelligence‐Driven Inverse Design of Singlet Fission Candidates in the Acene family. Advanced Science, Article e76828. https://doi.org/10.1002/advs.76828

Image Credits: AI Generated

DOI: 10.1002/advs.76828

Keywords: singlet fission, inverse design, machine learning, acenes, Hammett constants, GRU neural network, genetic algorithm, particle swarm optimization, photovoltaics, triplet excitons, DFT, chemical space

The commercial case for sodium-ion technology is strengthening rapidly. Recent industrial breakthroughs between 2025 and the present have seen announced manufacturing capacities exceed hundreds of gigawatt-hours, with particularly aggressive deployment in China. Commercial electric-vehicle packs now achieve energy densities above 170 watt-hours per kilogram, while the technology’s inherent safety, thermal stability and strong low-temperature performance make it ideal for grid-scale energy storage, where cost per kilowatt-hour matters more than compact size. High-power sodium chemistries are also attracting interest for drones, portable electronics and robotic platforms that demand rapid charging at low cost.

Yet sodium-ion batteries face stubborn material-level bottlenecks rooted in the element’s physics. The sodium ion is significantly larger than its lithium counterpart, with an ionic radius of 1.02 angstroms compared with 0.76 angstroms, leading to sluggish diffusion kinetics and substantial structural strain during repeated charging cycles. Layered oxide cathodes can undergo phase transitions and transition-metal migration that cause voltage fade, while hard carbon anodes suffer from low initial Coulombic efficiency and unstable interface formation. Alloy-type anodes based on tin, antimony and phosphorus promise theoretical capacities exceeding 1000 milliampere-hours per gram, but swell by more than 300 percent during sodiation, pulverizing particles and draining capacity. At high charging rates, non-uniform sodium deposition, dendrite growth and electrolyte decomposition compound these durability problems.

The review identifies four critical knowledge gaps that have kept traditional research methods from closing the performance gap with lithium. First, the solid electrolyte interphase, the protective film that forms on the anode, remains a molecular black box in sodium systems because sodium-based interphase components dissolve far more readily than their lithium analogues, driving continuous electrolyte consumption. Second, predicting high-voltage degradation and phase transitions with density functional theory calculations is computationally prohibitive. Third, high-quality standardized datasets for sodium materials lag dramatically behind the lithium literature. Fourth, many machine learning models function as inscrutable black boxes, offering predictions without the physical insight researchers need to understand structure-property relationships.

To overcome these obstacles, Hu argues, researchers are deploying an increasingly sophisticated AI toolbox. Graph neural networks represent crystal structures directly as atomic graphs, learning relationships between connectivity and properties without manual feature engineering, and have proven especially powerful for predicting formation energies, phase stability and electrode voltages in layered oxides. Deep neural networks serve as fast surrogate models for electrochemical performance metrics, while generative models such as generative adversarial networks, variational autoencoders and diffusion models can propose entirely new crystal structures that exist in no database. Inverse design flips the traditional workflow on its head: instead of making a material and measuring its properties, researchers specify desired performance targets and let the algorithm identify candidates likely to meet them.

Each cathode family demands a different AI strategy. For layered transition-metal oxides, machine learning models trained on density functional theory data distinguish P2 from O3 structures using descriptors like sodium concentration and mixing entropy, and AI-guided screening has identified phase-stable iron-nickel-manganese-titanium high-entropy oxides with improved cycling retention and air stability. Polyanionic cathodes of the sodium superionic conductor type benefit from active-learning frameworks that slash computational costs while exploring thousands of compositions, with generative models coupled to quantum simulations finding candidates with volume changes below 4 percent. Prussian blue analogues, whose modular open frameworks are plagued by vacancy defects and structural water, are ideally suited to variational autoencoders and Gaussian process regression that optimize composition and suppress degradation mechanisms.

On the anode side, machine learning is equally transformative. Ensemble methods such as XGBoost and random forests trained on biomass-derived datasets now predict reversible capacity and initial Coulombic efficiency from synthesis parameters, revealing how carbonization temperature tunes porosity and graphitic ordering in hard carbon. For alloy anodes, crystal-structure search algorithms combined with machine-learning-accelerated annealing identify thermodynamically stable sodium-alloy phases, while high-entropy alloy concepts redistribute sodiation-induced stress. Anode-free designs, in which sodium plates directly onto a bare current collector for maximum energy density, rely on AI to engineer artificial interphases and electrolyte additives that suppress dendrites, with engineered sodium-metal interfaces reported to reach Coulombic efficiencies approaching 99.96 percent under favorable conditions.

Electrolyte development may be where AI delivers its greatest leverage, because the design space of solvent, salt and additive combinations is astronomically large. Descriptor-based screening using HOMO-LUMO energies, dielectric constants and solvation energies allows algorithms to shortlist solvents that optimize sodium-ion coordination, while deep learning correlates additive chemistry with the formation of inorganic-rich protective layers composed of sodium fluoride, sodium oxide and sodium carbonate. Bayesian optimization combined with quantum calculations rapidly identifies additives that outperform conventional fluoroethylene carbonate, and physics-informed machine learning embeds thermodynamic constraints into models to accelerate prediction of ionic conductivity while preserving physical consistency. The same frameworks are being used to design localized high-concentration and fluorine-free electrolytes that suppress dendrite growth through favorable solvation structures.

The review’s most ambitious vision is the fully autonomous, closed-loop discovery platform, in which AI prediction, robotic synthesis and automated electrochemical testing feed results back into ever-improving models without human intervention. Active learning continuously selects the most informative experiments, Bayesian optimization explores sparse chemical spaces efficiently, and emerging agentic materials-science frameworks can plan, execute and refine entire development workflows. But Hu is candid about the remaining hurdles: sodium-specific datasets remain too small and scattered, dynamic interfacial phenomena span too many length and time scales for current models, and explainable AI methods will be essential to convert black-box predictions into actionable scientific insight. If those challenges can be met, self-driving laboratories could compress the journey from computational concept to validated battery material from years to months, accelerating sodium-ion batteries toward their role as the sustainable backbone of grid storage, electric mobility and beyond.

Subject of Research: AI-driven discovery of cathode, anode and electrolyte materials for sodium-ion batteries

Article Title: Review of AI-driven discovery of cathode anode and electrolyte materials for advancing sodium-ion batteries

Article References: Hu, Z. (2026). Review of AI-driven discovery of cathode anode and electrolyte materials for advancing sodium-ion batteries. Discover Industrial Chemistry and Materials, 1(1), Article 21. https://doi.org/10.1007/s44508-026-00022-x

Image Credits: AI Generated

DOI: 10.1007/s44508-026-00022-x

Keywords: sodium-ion batteries, artificial intelligence, machine learning, graph neural networks, generative AI, cathode materials, anode materials, electrolytes, solid electrolyte interphase, inverse design, high-throughput screening, autonomous laboratories

The study, published in the Journal of Materials Science: Metallurgy, describes what the authors call an inverse design methodology for brazed ceramic-metal composites. Brazing is the workhorse technique for joining ceramics to metals: the assembly is heated above the melting point of a filler alloy, which flows into the gap by capillary action and solidifies into a bond. The trouble is that ceramics and metals rarely share a coefficient of thermal expansion, or CTE. When the joint cools, that mismatch leaves residual stress locked into the interface, and repeated thermal cycling in orbit or during launch can turn those stresses into early failure. Traditionally, engineers would fabricate candidate joints, test them destructively or non-destructively, and iterate until they found a combination that survived.

That conventional route is punishingly slow. Destructive testing destroys specimens and generates waste; non-destructive methods such as X-ray computed tomography reveal defects but require costly equipment; and finite element simulations, while informative, take roughly twenty-four hours per joint assembly on the team’s setup. The Roorkee researchers, led by Sunita Khod and Mayank Goswami, asked whether a trained AI model could compress that entire loop. Their answer was to build a dataset linking material properties to joint performance, then train several machine learning and deep learning models to run the relationship in reverse: given a target average stress, the model outputs the thermal and elastic properties of the ceramic, braze, and metal that would produce it.

The foundation of the work is a carefully validated physics pipeline. The team modeled simple cube-based joint assemblies in the finite element software Abaqus, heating them from 300 to 1200 kelvin in vacuum and computing the average Von Mises stress, abbreviated VMS-avg, across the whole assembly. This average stress serves as a quantitative proxy for the joint’s load-bearing capacity, while peak local stress marks where failure would begin. To make sure the simulations meant something, the researchers compared them against published experimental shear strengths for two real brazed joints, alumina joined to niobium with a silver-copper-titanium filler, and zirconia joined to 316 stainless steel. The simulated stresses of 86.66 and 119.4 megapascals fell safely below the measured shear strengths of 150 and 300 megapascals, exactly what you would expect for joints that survive their thermal loading, confirming the physics was trustworthy.

Into the dataset went the coefficients of thermal expansion and Young’s moduli of candidate ceramics, brazes, and metals, drawn from the literature for materials including alumina, silicon nitride, Kovar, Monel, copper, and steel. The team also varied the porosity of the braze layer, modeling pores with volumes from 0.074 to 1.59 cubic millimeters and overall porosity between 0.003 and 0.49 percent, a range matching real space-industry components inspected by X-ray CT. Eighty-eight real data points emerged from the simulations. Because that is a modest number for machine learning, the researchers augmented it to 500 points using a random weighted linear interpolation with Dirichlet-distributed weights, a mathematical trick that generates new samples as convex combinations of real ones, guaranteeing every synthetic point stays physically plausible. Principal component and t-SNE analyses confirmed the augmented data preserved the structure of the original.

Then came the model bake-off. The team trained linear regression, polynomial regression, random forest, an artificial neural network, and a deep learning autoencoder, tuning each with cross-validation across eight different input-output configurations. A feature importance and SHAP analysis revealed that the stiffness of the metal and ceramic, their Young’s moduli, dominated the stress prediction, followed by their thermal expansion coefficients, while porosity and braze properties mattered less within the narrow porosity range modeled. On a held-out test set of 28 real, non-augmented points, polynomial regression and the neural network performed best for single-property predictions, achieving coefficients of determination near 0.95, while the random forest struggled badly on braze properties. But all four conventional models faltered when asked to predict multiple material properties simultaneously, a crucial capability for real design work.

That is where the autoencoder stole the show. An autoencoder learns to compress input data into a compact latent representation and then reconstruct it, and the team’s version, with a 32-16-32 hidden layer structure, proved remarkably good at capturing the tangled nonlinear relationships among thermal and mechanical properties. It achieved absolute percentage errors of roughly 0.125 to 4.5 percent against literature values on unseen data, and unlike the other models it needed to be trained only once, then queried for any output configuration by masking the relevant features. The authors argue the architecture has a deeper advantage: by restricting its latent space to physically realizable material combinations, it avoids the non-physical averaging that plagues direct regression, always returning a realistic ceramic-braze-metal trio rather than a chimera that exists only on paper.

The proof came in the predictions. Fed a target average stress of 87.08 megapascals, the model selected alumina ceramic, an Ag-Cu-Ti braze alloy, and Kovar, a low-expansion iron-nickel-cobalt alloy, a combination whose matched thermal expansion coefficients keep joint stress low. A higher target of 228.25 megapascals yielded silicon nitride with an InCuSil-ABA filler and Monel-400, a pairing with a much larger expansion mismatch and correspondingly higher internal stress, exactly as physics demands. These are not exotic picks; alumina, Ag-Cu-Ti, and Kovar are staples of aerospace feedthroughs, connectors, and insulators, and the same material families appear in travelling wave tube amplifiers and thermal protection systems. The model’s selections passed a physical sanity check, which is precisely what you want from an AI asked to make engineering decisions.

The practical payoff is speed. Once trained, the model returns candidate material properties in milliseconds, where a single finite element simulation takes about a day and a destructive test campaign takes weeks. An engineer could specify a safety threshold, say an average stress below 100 megapascals, and receive the required thermal and elastic properties of the braze, ceramic, and metal instantly, using the tool as a screening step before any fabrication begins. The authors are candid about the limits: the dataset is small, the model may not generalize to material classes far outside its training range, the simulations omit creep, viscoelasticity, and interfacial microstructures like intermetallic compounds, and the method is so far validated only for butt-joint geometries. Each simulation also demands serious computing power, which is why the team is already looking toward physics-informed neural networks and hybrid AI-finite element frameworks, along with larger experimental datasets and new joint geometries, to make inverse design a routine tool for the joints that hold spacecraft together.

Subject of Research: AI-driven inverse design for material selection in brazed ceramic-metal composite joints

Article Title: Artificial intelligence-driven methodology for predicting brazed ceramic–metal composite materials

Article References: Artificial intelligence-driven methodology for predicting brazed ceramic–metal composite materials. (n.d.). https://doi.org/10.1007/s44492-026-00004-2

Image Credits: AI Generated

DOI: 10.1007/s44492-026-00004-2

Keywords: artificial intelligence, brazing, ceramic-metal composites, autoencoder, finite element method, machine learning, thermal stress, aerospace materials, inverse design, materials selection, neural networks, coefficient of thermal expansion

Composite stiffened panels, the thin carbon-fiber skins reinforced by ribs and blades that make up much of a modern aircraft fuselage and wing, are deceptively complicated objects to specify. A single design involves decisions about the stiffener geometry, how many stiffeners to use, and two separate layup sequences, one for the skin and one for the stiffeners, all of which must simultaneously satisfy constraints on tensile stiffness, tensile load capacity, linear buckling load and post-buckling capacity. The number of feasible combinations explodes combinatorially, and each honest evaluation with finite element analysis is computationally expensive. The research team, led by authors from Shanghai Jiao Tong University together with collaborators at Northwestern Polytechnical University, the National Key Laboratory of Strength and Structural Integrity and the Aircraft Strength Research Institute of China, argues that this is precisely the setting where machine-learning surrogates earn their keep, provided they are deployed with enough statistical discipline.

The architecture of the new framework is deliberately split into three cooperating components. First, an inverse artificial neural network takes the required tensile stiffness, tensile load, linear buckling load and post-buckling capacity as inputs and generates multiple candidate designs, each described by geometry and layup variables. Second, a forward neural network, specifically weighted to evaluate buckling performance, scores each candidate against the original requirements. Third, a genetic algorithm performs a constrained, weight-minimizing search over the candidate space, but only after an engineering repair step has filtered out designs that violate manufacturing rules for composite laminates, such as balanced and symmetric stacking conventions. The entire pipeline is trained on data generated by parametric finite element analysis, so the surrogates learn from physics simulations rather than from experiments that would be impractical to run at scale.

The reported accuracy figures are notable for a surrogate model covering so many coupled responses. The forward neural network achieved an overall test-set mean absolute percentage error of 7.35 percent, with every individual response remaining below 10 percent error. In practical terms, this means the network can predict how a given panel design will behave under tension and buckling closely enough to guide optimization, while remaining fast enough to evaluate thousands of candidates in the time a single finite element solve might take. The authors are careful, however, not to oversell what the surrogate can do, a caution that runs through the entire paper and arguably constitutes its most important message.

That caution becomes concrete in the validation statistics. Across 50 distinct test requirements, with 50 candidate designs generated for each requirement, 91.52 percent of all candidates were judged feasible by the surrogate model. More importantly for real engineering use, every single requirement had at least one surrogate-feasible candidate among its top five suggestions, meaning the inverse network reliably produces usable options rather than occasional lucky hits. The team then ran a stratified optimization campaign: 20 requirements combined with five different genetic algorithm initialization strategies yielded 100 independent optimization runs, and all 100 runs converged to surrogate-feasible optima. The best-performing initialization scheme, designated the Hybrid strategy, reduced a weight proxy by an average of 14.11 percent, a figure that translates directly into the kind of mass savings that aircraft manufacturers chase relentlessly, since every kilogram removed from a primary structure compounds across thousands of panels over an airframe’s lifetime.

Yet the study’s most rigorous section concerns what happened after the optimization. Fifteen candidates with high safety margins were submitted to full Abaqus finite element back-checking, the industry-standard high-fidelity verification step. Thirteen of those analyses completed successfully, and every completed case satisfied all four reported performance requirements. The two incomplete analyses serve as a quiet reminder that surrogate predictions, however accurate on average, live in a statistical world that finite element verification must ultimately adjudicate. The authors draw a deliberate and somewhat contrarian conclusion from this: the inverse neural network should be used as a requirement-conditioned population initializer, seeding the genetic algorithm with intelligent starting points, rather than being trusted as a unique inverse solver that maps requirements to a single definitive design.

This framing matters because inverse design in engineering has often been marketed as a kind of oracle, a model into which you feed desired properties and out of which emerges the answer. The molecular and materials sciences have embraced generative inverse design with considerable enthusiasm, and structural engineering has followed. But the Chinese team’s results suggest a humbler and arguably more robust role: the inverse model excels at generating diverse, plausible starting candidates that would take conventional forward search enormous effort to discover, while the final word on feasibility and safety still belongs to physics-based verification. It is a hybrid philosophy that treats machine learning as an accelerator of the design funnel rather than a replacement for it.

The work also situates itself within a decade of rapid progress in surrogate-assisted composite optimization. Previous studies have used neural networks to optimize stacking sequences, genetic algorithms to minimize the weight of compressively loaded panels, Kriging models for post-buckling reliability analysis, and decision-tree methods to optimize laminate layouts for buckling resistance and imperfection sensitivity. More recently, researchers have applied neural-network-and-genetic-algorithm pairings to everything from pressure vessels to double-double composite structures, and physics-informed neural networks to the buckling of thin-walled cylinders. What distinguishes the new framework is its explicit insistence on conditioning the entire design process on performance requirements from the very first step, and on quantifying not just average accuracy but the probability that any given requirement will receive a feasible candidate at all.

The implications reach beyond a single panel type. As aircraft manufacturers push toward higher proportions of composite structure, the bottleneck is increasingly the engineering hours consumed by sizing, layup selection and verification cycles rather than raw material cost. A framework that can take a requirement table from a loads group and return dozens of near-optimal, repair-compliant, weight-minimized panel designs, each pre-screened for safety margin and queued for finite element confirmation, compresses that cycle substantially. The authors’ recommendation to retain safety-margin screening and finite element back-checking before engineering acceptance is not a caveat bolted onto the conclusion; it is the operating principle that makes the statistical results trustworthy in the first place. For an industry where a single undetected buckling mode can ground a fleet, the message is that artificial intelligence belongs in the driver’s seat of design exploration, with finite element analysis keeping its hand firmly on the brake.

Subject of Research: Machine-learning inverse design of composite stiffened aircraft panels using neural networks, genetic algorithms and finite element verification.

Article Title: Performance-driven ANN–GA inverse design and finite element back-checking of composite stiffened panels

Article References: Wang, Y., Luo, L., Yuan, M., Zhao, H., Chen, J., Wan, X., Chang, L., & Chen, J. (2026). Performance-driven ANN–GA inverse design and finite element back-checking of composite stiffened panels. Aerospace Systems. https://doi.org/10.1007/s42401-026-00538-2

Image Credits: AI Generated

DOI: 10.1007/s42401-026-00538-2

Keywords: composite stiffened panel, inverse design, artificial neural network, genetic algorithm, surrogate optimization, finite element analysis, buckling, post-buckling, layup sequence, lightweight design, aerospace structures, machine learning

The central achievement of the work is reconfigurability. Non-Abelian braiding, in which the outcome of swapping two objects depends on the order in which the swaps are performed, has previously been demonstrated in superconducting processors, trapped-ion systems, photonic chips and acoustic metamaterials. Those platforms are powerful but require cryogenic temperatures, intricate nanofabrication, or fixed on-chip geometries. The new experiment brings the same mathematical structure into an ordinary liquid crystal cell sitting at ambient conditions, where the relevant objects are micron-scale defect lines entangled around colloidal particles, and where the control knob is simply light.

Disclination lines are the skeletons of disorder within an ordered medium. In a nematic liquid crystal, rod-like molecules align with a local direction called the director; a disclination line marks a seam where that alignment field cannot be smoothly defined, much like the seam on a tennis ball marks a place where the covering cannot lie flat. When colloidal particles are dispersed in the nematic, their surfaces impose orientation constraints on the surrounding director, and defect lines become entangled among the particles, forming stable, topologically protected structures. The team exploited this entanglement by photonically manipulating the colloids, using patterned light to reorient the director field at the cell surface and drive the particles through cooperative molecular reorientations.

By sweeping this optical control, the researchers wove disclination lines into chiral double-helix entanglements, structures in which pairs of defect lines wind around one another in either a left-handed or right-handed twist. The handedness of the helix is not merely decorative; it serves as a binary degree of freedom that the team calls a nematic bit, or nbit. Depending on whether the surface director is rotated counterclockwise or clockwise by the incident light, the entanglement settles into one or the other chiral state, effectively writing a bit into the topology of the defect network. Supplementary videos accompanying the paper show the process unfolding in real time, with double-helix entanglements forming around assemblies of up to nine colloids and even coexisting regions of opposite chirality within a single four-particle structure.

With this encoding in hand, the team implemented a complete set of braid operations, the elementary moves in which defect lines pass over and under one another, and demonstrated their non-commutativity in networks of three lines. In an Abelian world, performing operation A and then operation B yields the same final configuration as performing B and then A. In the nematic platform, the order matters: two braid sequences that differ only in their ordering leave the network in topologically distinct states. This order-dependence is the defining signature of non-Abelian behaviour and the property that makes braided anyons attractive as a basis for fault-tolerant information processing, because the encoded state depends on the global history of exchanges rather than on any local measurement.

A crucial advantage of the soft-matter setting is that the braid gates themselves can be moved. The colloidal particles act as physical gates that pin and route the disclination lines, and by repositioning these particles with optical tweezers and light-driven transport, the researchers reprogrammed the braiding sequence in situ without rebuilding the sample. Small variations in colloid position or local line curvature leave the topological state unchanged, a robustness the team verified directly by perturbing the networks and observing that the encoded configuration survived. The method also extends beyond three-line demonstrations to multiline architectures, suggesting a path toward larger braiding networks assembled from the same elementary components.

Perhaps the most forward-looking contribution is the establishment of an inverse-design framework. Rather than working forward from operations to outcomes, the team developed an algebraic procedure that runs in reverse: given a desired topological transformation, the framework compiles it into prescribed spatial routing of the defect lines together with layer-by-layer phase corrections. This compiler-like capability mirrors how electronic design automation tools translate logic descriptions into circuit layouts, and it transforms the platform from a demonstration apparatus into a programmable one. The researchers note that the scalability of light-driven manipulation makes this design loop practical, since the same optical interface that writes a single bit can in principle address many.

The significance of the result lies in its bridging role. Topological information processing has been dominated by quantum proposals, where non-Abelian anyons would protect quantum states against local noise. Realizing the same braiding algebra in a classical, room-temperature material does not produce a quantum computer, but it provides a tangible, inexpensive laboratory in which non-Abelian logic can be studied, visualized and engineered. The authors position the system as a programmable classical platform for robust topological transformations, one that connects the soft-matter physics of liquid crystals with the emerging field of topological information processing. Because liquid crystals are already the workhorse of display technology, there is also a plausible engineering pathway: the optical and electro-optical toolkits for addressing nematic cells at high resolution are mature and commercially available.

The work also builds on a rich lineage. Knot-theoretic descriptions of nematic defects date back more than a decade, when theorists showed that disclination networks carry invariants analogous to braids and that rewiring operations among them can be classified. Experimentalists later demonstrated reconfigurable knots and links in chiral nematic colloids, and recent theoretical work proposed nematic bits and universal logic gates based on defect topology. What the new study adds is dynamics and control: the ability to actively drive the system through braid operations on demand, to verify non-commutativity experimentally, and to reprogram the network at will. Earlier light-driven studies from the same group had shown collective transport and reconfigurable assembly of nematic colloids and active transformations of disclination networks, providing the technical foundation for the present braiding results.

Looking ahead, the researchers suggest that the platform could serve as a testing ground for algorithms and error models relevant to topological computing, while also raising its own questions about how complex braid networks relax, hold information and fail. The combination of a mathematical structure once confined to abstract group theory with something as tangible as soap-like molecules and glass beads, manipulated by nothing more exotic than patterned light, is a reminder that some of the deepest ideas in physics can be made to run, quite literally, at room temperature. For now, the nematic bits weave their helices quietly under a microscope, but they weave them on command, in any order the operator chooses, and undo nothing by accident.

Subject of Research: Light-driven reconfigurable non-Abelian braiding of disclination lines in nematic liquid crystals

Article Title: Reconfigurable non-Abelian braiding of nematic bits

Article References: Lei, Z., Zheng, X., Zhang, J., Tang, W., Tian, K., Song, G., Asilehan, Z., Chen, Z., Vergara, F., Guan, Y., Zhang, R., Jiang, J., & Peng, C. (2026). Reconfigurable non-Abelian braiding of nematic bits. Nature Materials. https://doi.org/10.1038/s41563-026-02728-x

Image Credits: AI Generated

DOI: 10.1038/s41563-026-02728-x

Keywords: non-Abelian braiding, nematic liquid crystals, topological defects, disclination lines, nematic bits, colloids, topological information processing, soft matter, inverse design, liquid crystal colloids, chiral double helix, optical control