The problem the EPFL researchers, led by Mikael Reichler, Simone Bianconi and Tobias Kippenberg, set out to solve is rooted in the physics of optical feedback. In a conventional self-injection-locked system, a distributed-feedback (DFB) diode is coupled to a silicon nitride microresonator whose resonances act like an extremely selective mirror. Light reflected from the resonator pulls the laser frequency toward the resonance, and the laser’s noise collapses. But two conditions must be satisfied simultaneously: the free-running laser frequency must sit close enough to a cavity resonance, and the phase of the back-reflected light must fall within a narrow window. Temperature drifts, fabrication tolerances and the natural tuning of the laser with injection current all conspire to break these conditions. In practice, as the drive current is swept, the laser repeatedly falls out of lock into its noisy free-running state, and engineers must surround the chip with photodetectors and electronic feedback loops to hold it in place, inflating size, power consumption, cost and complexity.

The EPFL team’s insight was to identify a previously unrecognized control parameter that governs how the laser behaves as it tunes from one cavity resonance to the next. They call it feedback-phase dispersion: the difference in the phase of the back-reflected light between consecutive resonances of the microresonator. Because the resonator supports a comb of resonances separated by its free spectral range, and because the phase of the reflected light is periodic, the relative phase from one resonance to the next determines whether the laser, while tuning, always has a good locking state available or instead passes through regions where only poor locking exists. If the locking ranges of adjacent resonances overlap in the right way, the laser is forced to hop directly from one narrow-linewidth state to another, never touching the free-running regime.

Crucially, the researchers found that this dispersion can be engineered geometrically, and at a scale far coarser than the optical wavelength. The phase of light reflected from the resonator depends on the round-trip delays in the bus waveguide that feeds the resonator and the drop waveguide that extracts it, as well as on the position of the drop coupler along the resonator’s spiral path. A simple expression relates the feedback-phase dispersion to these delays and to the resonator’s free spectral range. For a typical 1-gigahertz free spectral range, the relevant length scale is about 30 centimetres, meaning the coupler positions need only be controlled to within centimetres rather than nanometres. That is a manufacturing-friendly tolerance, and it is what makes the approach scalable rather than a laboratory curiosity.

To make the feedback deterministic, the team equipped their spiral silicon nitride resonators with a drop port terminated by an integrated Sagnac loop mirror, which robustly excites the counterpropagating resonator mode and reflects it back to the laser. Without such a mirror, the feedback comes from random Rayleigh backscattering, and its phase fluctuates unpredictably from device to device. With the mirror, the phase evolution becomes stable and, more importantly, predictable from the layout. The researchers fabricated three resonator variants with different port placements and measured their light-current curves. The device without a drop port showed irregular, fluctuating resonance shapes, while the mirror-terminated devices showed stable, reproducible phase evolution that matched the design formula closely, confirming that feedback-phase dispersion is a genuine, engineerable quantity.

The demonstration of endless locking itself is striking. The team butt-coupled an off-the-shelf DFB diode emitting at 1,548 nanometres to a spiral resonator with a free spectral range of 1.44 gigahertz and an intrinsic quality factor of 11.3 million, engineered to have a positive feedback-phase dispersion of 0.16π. When the coupling between the diode and the chip was deliberately weakened, the light-current curve showed the familiar signature of failure: locked plateaus separated by bright intervals in which the laser fell free-running. When the coupling was raised to its design value, the plateaus merged, and the laser transitioned directly from one locked state to the next. Beating the output against a stable reference laser while sweeping the current over more than two full periods of the feedback-phase oscillation revealed a spectrogram in which the laser never once dropped into a free-running state, hopping instead across multiple resonances whenever doing so improved the linewidth.

The contrast case sealed the argument. A second resonator, nearly identical in every parameter except that its port placement produced a negative feedback-phase dispersion of −0.57π, showed exactly the predicted pathology: sections of the current sweep in which the laser unlocked and ran free. Numerical simulations reinforced the picture, showing that positive dispersion values allow the laser to select a high-quality locking state at every possible free-running frequency, whereas values near zero or negative leave gaps in which only weak locking exists. The team defines an endlessly locked laser quantitatively as one whose linewidth reduction factor exceeds 1,000 across the entire drive current range, and their device met that bar comfortably.

The measured performance places the chip among the best integrated lasers ever reported. Sweeping the injection current from 154 to 300 milliamperes in 2-milliampere steps, with no active control of current or phase, the researchers recorded frequency noise reductions of more than 5,000 relative to the free-running diode at every single step. Intrinsic linewidths fell below 10 hertz, a reduction of at least a factor of 10,000, approaching the fundamental thermorefractive noise floor of the silicon nitride resonator itself, the point at which tiny temperature fluctuations in the glass set the ultimate limit. The team also integrated monolithic piezoelectric actuators onto the chip, which strain the resonator to tune its resonances rapidly and with low hysteresis. Driving these actuators with a 10-kilohertz triangular waveform produced mode-hop-free frequency chirps spanning 1.5 gigahertz, exceeding the resonator’s free spectral range, at every injection current, with state-of-the-art linearity.

What elevates the result beyond a record is its generality. The authors show through simulations that the design rule, a relatively low free spectral range combined with a non-zero, preferably positive feedback-phase dispersion, should extend to Fabry–Pérot diodes, to resonators with free spectral ranges beyond 2 gigahertz when the dispersion is near 0.6π, and to photonic platforms operating from the ultraviolet through the mid-infrared. The underlying physics applies equally to bulk crystalline microresonators. With output power of several milliwatts, comparable to state-of-the-art precision lasers, and a footprint set by a single diode chip and a photonic integrated circuit, the architecture converts self-injection locking from a finicky laboratory procedure into a turnkey component. For applications from optical clocks and quantum sensors to coherent LiDAR and fibre sensing, that could mean hertz-level coherence no longer confined to optics tables, but manufactured by the millions on wafers and deployed wherever a quiet, stable beam of light is needed.

Subject of Research: Self-injection-locked silicon nitride photonic integrated lasers with engineered feedback-phase dispersion for robust ultralow-noise operation

Article Title: Endlessly self-injection-locked photonic integrated lasers

Article References: Reichler, M., Bianconi, S., Zhang, Y., Liffredo, M., Villanueva, L. G., & Kippenberg, T. J. (2026). Endlessly self-injection-locked photonic integrated lasers. Nature Photonics, 20(10), 1146-1152. https://doi.org/10.1038/s41566-026-01985-1

Image Credits: AI Generated

DOI: 10.1038/s41566-026-01985-1

Keywords: self-injection locking, photonic integrated circuits, silicon nitride, narrow-linewidth lasers, microresonators, semiconductor lasers, feedback phase, frequency noise, optical atomic clocks, LiDAR, piezoelectric tuning, EPFL

The study, published in Nature Microbiology by Alexandre Persat’s Microbial Mechanics Lab in collaboration with Sangwoo Kim’s Mechanics of Soft and Biological Matter Laboratory, both at EPFL, reveals that the opportunistic pathogen Pseudomonas aeruginosa actively senses when it collides with neighboring cells and responds by reversing its direction of travel within seconds. This touch-based feedback loop allows the bacterial population to do something remarkable: stay highly organized where organization matters, such as at the advancing edge of a growing colony, while deliberately dissolving that order in crowded regions where it would otherwise trap individual cells. According to the authors, this shows that collective order in bacterial groups is not merely a passive consequence of physics but an actively regulated state that the cells themselves control.

The problem the bacteria face is one that anyone who has ever been stuck in a dense crowd will recognize. P. aeruginosa moves across surfaces using a form of motility that depends on physical contact with the substrate, and in dense communities the cells frequently line up and stream together in coordinated groups. These physical interactions between neighboring cells create alignment, and alignment creates order. But order has a downside. Individual cells embedded within a coordinated stream can become trapped inside the group, unable to turn away and strike out into new territory. For a bacterium, that inability to disperse is a serious handicap: it prevents the cell from finding fresh nutrients, and it hinders escape from crowded, structurally complex environments such as soil pores or the tissues of a host during infection.

Laure Le Blanc, the paper’s first author, explains the mechanism in simple terms. As the bacteria move across surfaces, they sense when they bump into neighbors, and in response they reverse direction within seconds. That rapid reversal is the key. Each collision acts as a piece of sensory information, telling the cell that its local environment is congested and that continuing straight ahead would be counterproductive. By reversing, the cell injects disorder into the collective motion, loosening the alignment that physical contact tends to create. Where collisions are rare, for example at the open frontier of a colony, the cells align strongly and move together toward unexplored space. Where collisions are frequent, the reversals accumulate and the group remains disordered and mobile. The population effectively uses touch to tune its own degree of collective organization from moment to moment.

To disentangle how individual collisions translate into group-level behavior, the team combined three complementary approaches: live microscopy to watch the bacteria in action, single-cell tracking to follow the trajectories of individual cells, and computer simulations to test how collision-driven reversals shape population dynamics. Crucially, they compared wild-type bacteria, which possess the full sensory machinery, with mutant strains that lack it. This side-by-side comparison allowed the researchers to isolate the specific contribution of collision sensing, rather than inferring it indirectly from the behavior of intact populations alone. By tracking individuals and analyzing overall group movement simultaneously, they could observe in real time how a microscopic bump between two cells rippled outward to reshape the behavior of the entire community.

The contrast between the two strains was striking. The researchers worked with a mutant lacking a gene critical for the bacterium’s normal motility pattern; the mutant always moves forward, even after colliding with other cells or with boundaries. When traveling alone, the mutants actually performed well, covering greater distances than their wild-type counterparts, since relentless forward motion pays off in empty space. But when moving together, the advantage evaporated. Without the ability to respond to collisions, the cells quickly became trapped in dense clusters, jamming against one another much like a crowd pressing into a blocked exit. The wild-type cells, by contrast, used each bump as information, reversing in crowded areas to keep the group fluid and exploratory, while maintaining tight alignment at the colony edge where open space invited coordinated advance.

The most dramatic demonstration came from an experiment that sounds almost whimsical: the researchers built micro-mazes, tiny obstacle-filled labyrinths, and released the two strains inside to see how they would navigate the confined, structured environment. The results were decisive. Only one percent of the collision-blind mutant bacteria managed to reach the maze exit within four hours. The wild-type cells, equipped with their touch-based navigation system, explored an average of ninety-two percent of the maze, and twenty percent of the cells eventually found the exit. In other words, the ability to feel a collision and reverse course transformed a population of microscopic wanderers into efficient problem-solvers capable of threading their way through a physical puzzle that stumped the mutants almost completely.

At the mechanistic level, the findings reframe what bacterial collectives are doing on surfaces. The classical picture treats alignment and streaming in dense bacterial groups as emergent physics, the inevitable result of cells pushing against one another as they crawl. The new work shows that P. aeruginosa actively modulates that physics, converting mechanical interactions into sensory input that feeds back onto individual behavior and, from there, onto the architecture of the whole population. The cells are, in effect, performing a computation: the frequency of collisions in their immediate surroundings encodes the local density of the crowd, and the reversal response reads that signal out as a behavioral decision. This allows the same population to display two very different collective states, ordered at the frontier and disordered in the interior, without any central coordination or chemical signaling between distant cells.

The implications extend well beyond basic microbiology. P. aeruginosa is a notorious opportunistic pathogen, and the researchers suggest that this collision-sensing mechanism may help the bacterium spread through host tissues and contribute to biofilm formation during human infections. Understanding how the cells regulate their collective organization could therefore inform new strategies for interfering with bacterial dispersal and biofilm development, both of which are major challenges in clinical settings where antibiotic-resistant infections are a growing concern. The discovery also carries a lesson for engineers. Robot swarms deployed in crowded or unknown environments face exactly the same trade-off between cohesion and exploration that bacteria face, and they often get stuck for precisely the reasons the mutant cells do. A simple rule, reverse when you bump into something, turns out to be a remarkably effective decentralized algorithm for keeping a swarm mobile, and it requires no central controller, no map, and no communication beyond the moment of contact itself.

What makes the study especially compelling is the way it connects scales. A single molecular event, a cell brushing against a neighbor, triggers a reversal within seconds; thousands of such events sculpt the shape and exploratory capacity of an entire colony; and in a maze, that collective behavior determines whether the population escapes or perishes. Persat’s framing of the question, drawn from fish schools, bird flocks, and human crowds, finds its answer in one of the simplest organisms on Earth: individuals in a group can balance coordination with exploration by using their own bodies as sensors. For bacteria, the crowded environment is not an obstacle to navigation but the very source of the information needed to navigate it. The bump in the crowd, it turns out, is the message.

Subject of Research: Mechanosensing of cell-cell collisions regulating collective navigation in Pseudomonas aeruginosa

Article Title: Bacteria use collisions to navigate crowded environments

Article References: Bacteria use collisions to navigate crowded environments. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: Pseudomonas aeruginosa, bacteria, mechanosensing, collective behavior, cell-cell collisions, twitching motility, biofilms, micro-mazes, EPFL, Nature Microbiology, robot swarms, bacterial navigation

The task the researchers tackle, known as time series extrinsic regression, involves predicting a continuous target variable from an input time series. It underpins vital-sign forecasting in healthcare, volatility prediction in finance, pollution estimation in environmental monitoring, and remaining-useful-life estimation in engineering. In all of these settings, accuracy alone is not enough. If a model warns of an impending system failure, engineers need to trace that warning back to specific sensor behaviors and specific time intervals. Yet the strongest current models, from deep convolutional networks to ROCKET-style random kernel transforms and large ensembles, achieve their performance precisely by hiding their reasoning in millions of parameters.

The dominant workaround has been post-hoc explanation: training an opaque model first, then applying saliency maps or feature attribution methods such as Integrated Gradients or DeepLIFT to rationalize its decisions afterwards. The EPFL team argues that this approach is fundamentally fragile. Post-hoc explanations are often coarse, noisy, or unstable, and they are not guaranteed to reflect the model’s true internal reasoning. Different attribution methods can disagree with one another, and because interpretability constraints are imposed only after training, the explanations may bear little relation to how the prediction was actually computed. Meanwhile, inherently interpretable alternatives such as Neural Additive Models and their time-series extension NATMs decompose predictions into per-feature contributions but remain fundamentally univariate, unable to capture interactions between variables. Concept-based models offer richer structure but typically require concept annotations that rarely exist for time series data.

MAGNETS resolves this impasse by embedding transparency directly into the architecture while still allowing a powerful neural network to do the heavy lifting where it cannot corrupt interpretability. The pipeline unfolds in four stages. First, a mask generation network, implemented as a one-dimensional U-Net, examines the input time series and proposes a set of binary masks, one per channel and mask index, that select the temporal regions relevant to the prediction. Second, each mask is applied to the raw input by element-wise multiplication, zeroing out everything outside the selected regions. Third, the masked values are aggregated over time into a single scalar per channel-mask pair, using a simple summation that captures both the duration of the relevant pattern and its intensity. Finally, these aggregated features pass through a linear concept bottleneck and a strictly linear prediction layer, so every output can be traced back, step by transparent step, to specific parts of the original signal.

The technical heart of the method lies in how the binary masks are trained. Because hard, all-or-nothing masks would normally block gradient flow, the authors employ the Straight-Through Gumbel-Softmax estimator: during the forward pass the masks are binarized so that the selected regions are crisp and unambiguous, while during the backward pass gradients flow through a relaxed, differentiable version. The team deliberately chose binary rather than soft masks, arguing that in continuous time series, where the precise magnitude and shape of values matter, blurred probabilistic masks would obscure exactly which parts of the signal drive the prediction. The masks are also input-specific rather than global, meaning the model can point to different relevant intervals for different individual samples, and a single mask may select several disjoint intervals within a channel.

Two regularization terms keep the learned concepts compact and non-redundant. A sparsity penalty, based on the L1 norm of the bottleneck weights, encourages each concept to depend on only a small subset of the aggregated features. An orthogonality loss, computed as the squared Frobenius norm of the deviation of the weight Gram matrix from the identity, discourages different concepts from encoding similar combinations of features. Together, these constraints produce concepts that are both sparse and distinct, making it clear which masked temporal regions and which channels drive each concept activation. The final prediction is a weighted linear combination of the concept activations, so each concept contributes additively and with an explicit, inspectable weight.

To test whether the model actually recovers the right reasoning, the researchers built four synthetic datasets with known ground-truth logic, ranging from a simple univariate rule, where the target equals the area under the curve above a threshold, to a challenging trivariate task where the target is a weighted sum of three conditional areas defined by pairwise comparisons between channels. The results were striking. MAGNETS outperformed all black-box baselines, often by a substantial margin, on the Bivariate and Trivariate-2 tasks, and beat every existing interpretable model across all multivariate settings. On the hardest trivariate dataset, the regularized bottleneck learned three disentangled concepts that corresponded directly to the correct channel relationships, including one concept that depended jointly on two channels, reconstructing the underlying symbolic logic without any supervision. By contrast, DeepLIFT attributions for a black-box CNN fluctuated sharply across neighboring time steps and failed to capture the necessary multivariate relations.

The evaluation extended to ten real-world datasets spanning household power consumption, wind turbine power, benzene concentration, Beijing air quality, flood modeling, battery degradation, and a newly released bridge degradation benchmark. Here, unconstrained black-box models generally retained the lowest error, as expected, but MAGNETS narrowed the gap dramatically and even surpassed all baselines on FloodModeling2 and BatteryDegradation1. Among interpretable models, it was the consistent leader, achieving the lowest root mean square error on seven of the ten datasets while remaining highly competitive on the rest. Notably, the team applied a single fixed configuration across all tasks, with three concepts and ten masks per channel, and a sensitivity analysis confirmed that performance remained stable across a wide range of hyperparameters, demonstrating that the architecture does not depend on per-dataset tuning.

The qualitative analyses may prove the most compelling for practitioners. On the bridge degradation dataset, MAGNETS discovered a meaningful dependency between bridge displacement and train load, consistently isolating intervals where train load exceeded roughly 2.0 and highlighting the corresponding displacement values. This supports a concrete, testable hypothesis: that structural degradation depends primarily on displacement during intervals of critical loading. On the battery dataset, the model focused on the initial voltage drop of a constant-current discharge, a known indicator of internal resistance growth, and the tail of an open-circuit voltage discharge, which reflects remaining active material capacity. These physically meaningful patterns emerged automatically, aligned with battery chemistry, and without any concept annotations.

The work directly confronts the long-standing performance-interpretability trade-off, showing that for regression tasks driven by discrete, localized temporal events, transparency need not come at the cost of accuracy, and can even enhance it. The authors acknowledge limitations that open future directions: real-world explanation correctness still lacks quantitative ground truth, the current aggregation is restricted to summation, and extending the framework to time series classification is a natural next step. But the core message is clear and likely to resonate far beyond machine learning research. By learning to mask and aggregate, MAGNETS answers the three questions its title poses, when a pattern occurs, how long it lasts, and how much it contributes, in a form that domain experts can actually inspect, validate, and act upon. In safety-critical domains from healthcare to civil infrastructure, that may matter more than another decimal point of accuracy.

Subject of Research: Inherently interpretable neural networks for time series extrinsic regression using learned mask-based concept aggregation

Article Title: When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate

Article References: Forest, F., Wei, A., & Fink, O. (2026). When, how long and how much? Interpretable neural networks for time series regression by learning to mask and aggregate. Data Mining and Knowledge Discovery, 40(6), Article 103. https://doi.org/10.1007/s10618-026-01267-y

Image Credits: AI Generated

DOI: 10.1007/s10618-026-01267-y

Keywords: time series regression, interpretable machine learning, explainable AI, concept bottleneck models, neural networks, mask learning, EPFL, battery degradation, structural health monitoring, Data Mining and Knowledge Discovery, U-Net, Gumbel-Softmax