A team led by Tianzhong Huang of the Shanghai Marine Equipment Research Institute, together with colleagues from Jiangsu University of Science Technology and Zhejiang College of Security Technology, tackled a 20-tonne-class wave compensation platform whose geometry had never been systematically optimized. Their approach, detailed in a paper published on 28 July 2026, integrates an artificial neural network surrogate model with an enhanced version of the NSGA-III multi-objective genetic algorithm, followed by TOPSIS decision analysis to select a single winning configuration. The result is a design that improves dexterity by 18.12 percent, cuts peak driving force by 11.78 percent, and enhances the kinematic condition number by 6.85 percent, all while slightly reducing actuator stroke.

The engineering stakes are considerable. In a Stewart platform, six hydraulic or electromechanical legs connect a base to a moving platform, and the radii of the two joint circles, the angular distribution of the joints, and the initial platform height jointly determine how the machine behaves across its entire workspace. Get these dimensions wrong, the authors note, and the consequences range from wasted deck space and redundant actuator strokes to legs that endure dangerous force spikes during extreme wave poses. Worse still, if the mechanism drifts toward a singular configuration, the platform can oscillate violently, risking mechanical failure at precisely the moment reliability matters most.

Previous research on wave compensation has concentrated heavily on control algorithms, from linear active disturbance rejection and neural-network-based PID schemes to model predictive control and linear-quadratic-Gaussian methods. But as the researchers point out, most of these strategies assume fixed mechanism dimensions, using software to compensate for physical design deficiencies. Any control scheme, however sophisticated, remains constrained by inherent physical bottlenecks such as actuator output limits and joint interference. The new work flips the emphasis, arguing that a rational physical configuration must come first, with control layered on top of an already well-conditioned structure.

The optimization problem itself is genuinely multi-dimensional and riddled with trade-offs. The team defined a five-dimensional decision vector comprising the upper and lower platform joint radii, the two joint distribution angles, and the initial height, then formulated four competing objectives: minimizing maximum actuator stroke, minimizing peak driving force, maximizing global dexterity, and minimizing the kinematic condition number, which measures how uniformly the platform responds across all six degrees of freedom. These indices are mutually exclusive in practice, and earlier attempts to blend them with weighted single-objective methods suffered from subjective weight choices and premature convergence in the face of the Stewart platform’s notoriously nonlinear, multimodal design landscape.

Computational cost posed another barrier. Evaluating peak driving forces requires traversing extreme poses and solving force Jacobian equations derived from the principle of virtual work, while dexterity and isotropy demand singular value decomposition of the Jacobian matrix at every pose. To break this bottleneck, the researchers generated 2500 datasets with a MATLAB-based physical simulation program, using random sampling within engineering-constrained boundaries, and trained a four-layer neural network with a 5-20-15-4 architecture and sigmoid activations. Trained via the Levenberg-Marquardt algorithm on an 80:20 train-test split, the surrogate achieved average R-squared values above 0.98 across all four objectives. Even when 0.5 percent Gaussian noise was injected to mimic manufacturing errors, joint clearances, and sensor noise, testing accuracy stayed between 0.987 and 0.999, an order-of-magnitude leap in optimization efficiency without sacrificing fidelity.

The optimization engine is where the study makes its most distinctive contributions. Standard NSGA-II loses its screening power as objectives multiply, so the team adopted NSGA-III, which steers evolution using 455 uniformly distributed reference points on a hyperplane in the four-dimensional objective space. Two enhancements proved decisive: a dynamic objective-space normalization strategy that eliminates the extreme scale bias between stroke metrics on the order of meters and force metrics on the order of hundreds of kilonewtons, and a constraint-violation penalty that explicitly forbids solutions violating cylinder stroke limits or exceeding the allowable cone angle of the Hooke’s joints. Head-to-head comparisons showed the improved NSGA-III reduced the spacing metric by 37.4 percent and boosted hypervolume by 26.2 percent relative to NSGA-II, producing a denser, more evenly distributed Pareto front of genuinely feasible designs.

Sensitivity analysis revealed that the upper platform radius and the initial height are the dominant levers of performance. Raising the upper radius from 2140 to 3110 millimeters, for instance, dropped peak driving force from 117.47 to 87.44 kilonewtons while correlating strongly with improved dexterity and a condition number falling from roughly 22 to 14.28. From the 455 non-dominated solutions, the team distilled three archetypal design families, a compact scheme for cramped decks, a robust scheme for heavy loads, and a dexterous scheme for kinematic performance, then applied TOPSIS with engineering-informed weights, giving peak driving force the highest weight at 0.30, to rank 91 shortlisted candidates. The top-ranked configuration expanded the base radius to 3881 millimeters and the upper radius to 3060 millimeters at a height of 5291 millimeters, with joint angles of 19.82 and 19.04 degrees that preserve safe installation clearances.

Critically, the team did not stop at simulation. They built a scaled proof-of-concept prototype with DC linear actuators, mirroring the optimized joint angles, and commanded it through roll and pitch extremes of plus and minus 15 degrees fifty consecutive times. The relative error between commanded and executed angles stayed strictly within 1 percent across every cycle, and the six legs operated smoothly with no collisions or interference between kinematic pairs. The authors are candid about limitations: the dynamic evaluation rests on rigid-body assumptions, so structural elasticity under extreme loads on the full 20-tonne system remains unmodeled, and the optimization deliberately excludes control co-design. Future work will pair multibody dynamics models with control-structure co-design to test real-time tracking under stochastic wave spectra.

For an offshore industry racing to install and maintain turbines in ever rougher seas, the study offers something rarer than a better gangway: a reusable design methodology. By fusing neural-network surrogates, reference-point-based evolutionary search, and principled multi-criteria decision-making, it demonstrates that the physics of a machine can be optimized before its software ever runs, providing a benchmark that could ripple through shipborne robotics, deployable antennas, and any heavy parallel mechanism where geometry decides survival.

Subject of Research: Multi-objective optimization of a heavy-duty wave compensation parallel platform using an ANN surrogate model and an improved NSGA-III algorithm

Article Title: Optimal design and validation of a heavy-duty wave compensation parallel platform utilizing the NSGA-III algorithm

Article References: Huang, T., Luo, R., Mao, L., Wang, J., Zhou, F., & Zhang, G. (2026). Optimal design and validation of a heavy-duty wave compensation parallel platform utilizing the NSGA-III algorithm. Mechanical Sciences, 17(2), 783-798. https://doi.org/10.5194/ms-17-783-2026

Image Credits: AI Generated

DOI: 10.5194/ms-17-783-2026

Keywords: wave compensation, Stewart platform, parallel mechanism, NSGA-III, artificial neural network, multi-objective optimization, TOPSIS, offshore gangway, dexterity, condition number, surrogate model, marine engineering

At the heart of the controversy is a technique the authors call SLRI, or spectral label-free Raman imaging, which combines spontaneous confocal hyperspectral Raman microscopy with a deterministic mathematical fitting procedure known as non-negative least squares, or NNLS. The method scans tissue point by point with a single 532-nanometer excitation beam, acquiring a full Raman spectrum at every pixel across thousands of spatial positions and multiple depth planes. Each measured spectrum is then fitted against a library of 42 purified reference compounds, ranging from saturated and unsaturated fatty acids to cytochrome c and other biomolecules, producing spatial maps that show where each reference component contributes most strongly to the measured signal.

The critical letter, according to the authors, misread the scope of these claims. The reply is unambiguous on this point: the published article never asserted that the NNLS-derived maps represent absolute molecular concentrations, nor that all 42 analytes had been definitively and independently identified. Instead, the paper explicitly describes the maps as reference-constrained relative distributions, and the small-molecule images as Raman-inferred, model-constrained representations. The authors lay out a clear analytical hierarchy to make the distinction concrete: a measured Raman spectrum feeds into a deterministic, reference-constrained NNLS coefficient, which in turn generates a spatial coefficient map. No step in the Drosophila study converts those coefficients into milligrams per milliliter or moles per liter. A separate concentration-calibrated implementation exists in the team’s earlier qRamanomics work, published in Cell Reports Methods in 2023, but that calibration was claimed only for selected major biomolecular classes, not for all 42 references.

Where the published text used shorthand such as a component being enriched in a region, the authors explain, the intended meaning is a relatively higher distribution of the corresponding reference-constrained spectral contribution, not a calibrated concentration increase. They also address a related criticism about statistics: violin plots in the supplementary figures quantified only three representative components, unsaturated fatty acids, saturated fatty acids, and oxidized cytochrome c, at the whole-testis level. These were chosen as illustrative examples of signal trends, the authors say, not because the imaging was limited to three molecules. Crucially, they acknowledge that pixels within a single specimen are spatially correlated measurements and were never treated as independent biological replicates, so pixel-level statistical tests were not used as evidence of biological replication.

On the question of whether the computational workflow was reconstructible from the public record, the authors concede the Advanced Science article could have linked the code and study-specific data more explicitly, but they reject the charge that the pipeline was opaque. The same MATLAB qRamanomics processing and NNLS codebase used in the study was previously published and archived on Zenodo, and the principal new analytical input was the study-specific 42-reference library. To close the loop, the team has now released a set of files that make the data lineage explicit: the normalized 42 reference spectra, a representative raw hyperspectral Z-plane containing 30,771 measured spectra on a 263 by 117 grid with 1,114 Raman-shift channels per spectrum, a numerical NNLS coefficient vector for a cytochrome c map, and the corresponding rendered TIFF image. Reshaping the coefficient vector to a 600 by 686 array reproduces the published image with a Pearson correlation of approximately 0.9999, a striking demonstration that the colorful figures trace directly back to raw pixel-resolved spectral measurements rather than to stylized graphics.

Perhaps the most technically substantive portion of the reply concerns identifiability, the question of whether 42 overlapping spectra can be reliably separated at all. The authors report new diagnostics on the exact reference library after interpolation onto the measured Raman axis and application of the fitting window from 400 to 3100 inverse centimeters, excluding the spectroscopically silent region between 1800 and 2700. The 42-by-42 design matrix is full rank, meaning no reference spectrum is an exact linear combination of the others, with 500 fitting variables available for 42 components. Yet the team does not oversell this result. The 2-norm condition number of the unit-area-normalized matrix is 391.5, dropping to 242.6 after additional L2 column normalization, and pairwise correlation and hierarchical clustering analyses reveal substantial collinearity among several chemically related references. The authors explicitly state that these diagnostics are not proof that all 42 components are mutually orthogonal or unconditionally identifiable; rather, they quantify the model dependence that motivated the cautious Raman-inferred terminology in the original paper. They also dismiss a tempting shortcut: refitting a noise-free synthetic mixture with the same design matrix would be mathematically circular and therefore meaningless as validation.

The reply also clarifies sample preparation, which bears directly on whether long acquisitions could have been corrupted by biological motion. The Raman-imaged testes were fixed in 4 percent paraformaldehyde for 20 minutes, washed three times in phosphate-buffered saline, embedded in agarose, and maintained in PBS during acquisition. The specimens were therefore not living tissues undergoing cell migration, differentiation, or metabolite turnover during scans that can nominally run for many hours. The authors further explain that the punctate and ring-like patterns highlighted in the study were identified directly within individual XY planes, comparing measured spectra from peripheral punctate loci and central diffuse loci within the same structure, so variation in z-step size cannot generate the observed patterns. Three-dimensional renderings simply stack consecutively acquired planes recorded in the same microscope coordinate system; no intensity-based registration between separately acquired channels is involved, because every molecular coefficient at a given coordinate derives from the same measured spectrum.

On the modality question, the authors note there is no experimental ambiguity: the study describes spontaneous confocal Raman microscopy with single-beam 532-nanometer excitation, point scanning, full-spectrum acquisition, and EMCCD detection, features that distinguish it from stimulated Raman scattering. The critical letter itself, they observe, reached the same classification, so the objection reduced to a citation-placement issue rather than a claim that the wrong technique was used. The authors also take aim at the conceptual figure accompanying the letter, pointing out that its microscope illustration corresponds to no confocal Raman instrument, its spectral traces are schematic drawings rather than experimentally acquired spectra, and its laser beam is drawn blue when 532-nanometer light is green. If AI-assisted tools generated the figure, they add, appropriate disclosure under journal policies would be important.

The broader stakes of the exchange concern the role of artificial intelligence in spectroscopy. The letter proposed an AI-ready multimodal spatial-biology framework as an alternative, but the authors argue that a conceptual schematic is not evidence of analytical superiority. AI may prove valuable for Raman denoising, representation learning, and multimodal integration, they concede, but it is neither necessary nor evidentiary for validating a deterministic study whose reproducibility rests on fixed inputs, a specified reference matrix, documented preprocessing, and NNLS optimization, all of which yield identical numerical outputs given identical inputs and code. They challenge the correspondent to present one or two core experimental datasets to substantiate the proposed framework.

The authors close by situating the Raman maps within a wider web of orthogonal evidence: two independent Cyp4ae1 RNAi lines consistently produced the spermatogonial differentiation arrest phenotype, while LD540 and TOM20 staining, ATP and NADP+/NADPH assays, and non-targeted metabolomics independently documented lipid accumulation and mitochondrial abnormalities. Bulk assays do not prove microscopic localization of any individual Raman component, the authors acknowledge, but they support the broader metabolic remodeling within which the spatial Raman-inferred patterns are embedded. With the full datasets now posted publicly, including the raw hyperspectral layers, reference library, and diagnostics, the team argues that the study rests on real pixel-resolved data and a documented, deterministic pipeline, and that the appropriate standard for evaluating any spectral imaging claim is a minimum evidentiary framework in which the strength of the interpretation never exceeds the strength of the underlying spectra.

Subject of Research: Label-free Raman imaging methodology and evidentiary standards for spectral map interpretation in biomedical research

Article Title: From Spectral Maps to Biomedical Evidence: A Minimum Evidentiary Framework for Label‐Free Raman Imaging (Reply to Letter‐to‐the‐Editor)

Article References: Li, J., Zhao, X., & Yu, J. (2026). From Spectral Maps to Biomedical Evidence: A Minimum Evidentiary Framework for Label‐Free Raman Imaging (Reply to Letter‐to‐the‐Editor). Advanced Science, Article e77930. https://doi.org/10.1002/advs.77930

Image Credits: AI Generated

DOI: 10.1002/advs.77930

Keywords: Raman spectroscopy, NNLS, label-free imaging, Drosophila, spectral unmixing, reproducibility, qRamanomics, hyperspectral imaging, condition number, biomedical evidence, Cyp4ae1, spatial biology