In the inertial confinement fusion (ICF) community, AI-assisted design is highly anticipated but currently relies on a fragmented toolkit: surrogate forward models, inverse Bayesian inference tools, automated design optimizations, etc., each trained for a single specific direction of information flow. Yet all these tasks share a common mathematical structure—they query conditional distributions of the same underlying joint distribution p(design, performance). We propose to learn this joint distribution directly. Design parameters and performance metrics are treated as tokens, and a transformer is trained with random-order masking to approximate their joint probability density. Once trained, the diverse tasks mentioned above all reduce to a single operation: evaluating a conditional probability under the learned distribution. Applied to a dataset of 81 920 one-dimensional implosion simulations, the model reproduces well-established ICF scaling laws—validating that it has internalized the correct physical constraints—and proposes new designs under up to eight stringent constraints with a high success rate, high diversity, and a ten-thousand-fold efficiency gain over brute-force screening. In high-dimensional sub-regions of the design space where only a handful of training samples survive, the model still returns the correct and well-normalized conditional distributions, demonstrating that it captures the underlying manifold structure of physical constraints rather than merely interpolating between data points. Our framework features a simple architecture and a firm probabilistic theoretic foundation. It provides a new approach for analyzing strongly coupled systems like the ICF implosion.
We present EUV-OVL-SYN, an open and fully reproducible synthetic benchmark dataset for EUV lithography. Overlay error, the misalignment between consecutive patterning layers, is a primary yield-limiting factor at advanced nodes, yet the absence of publicly available, physically realistic datasets impedes systematic progress in data-driven correction research. EUV-OVL-SYN is built on documented scanner physics and replicates the signal hierarchy of real production data. The dataset captures three distinct levels of physical structure: a dominant inter-field Brink/van den Brink polynomial fingerprint (cross-wafer R2 ≥ 0.93); four lot-level temporal drift scenarios modeling linear scanner heating, reticle alignment drift, multi-parameter process drift, and chiller-cycling oscillation; and a structured nonlinear intra-field residual analytically verified to lie outside the Brink polynomial basis. The dataset consists of 4 lots of 50 wafers each, measured at 420 metrology targets per wafer (284 000 measurements in total), formatted identically to industrial automatic process control exports for direct compatibility with existing correction pipelines. The nonlinear residual (1σ ≈ 1.1–1.4 nm per axis) is well above the 0.15 nm measurement noise floor yet non-trivially masked by the dominant Brink fingerprint, providing a detectable but genuinely challenging correction target. Full ground-truth decomposition is provided for every signal component, permitting exact, method-agnostic evaluation of any correction approach. The complete dataset is generated by a single reproducible Python script and released under an open license.
Advancements in laser-based secondary radiation sources are being made by controlling the pulse shape of ultrafast, high-power laser systems away from its best-compressed level (shortest duration). Pulse shaping at the femtosecond level relies on control over a combination of both the laser spectral intensity and phase. Devices such as an acousto-optic programmable dispersive filter are often used at high power (>1 TW) laser facilities to affect the spectral phase while largely preserving the spectral intensity and overall laser power. Using this type of controllable hardware, multiple machine-learning-based models were trained on experimental data. This was performed to capture all nonlinearities in the amplification process and to predict upstream laser input parameters needed to produce a specific output pulse shape at TW-level peak powers. Comparisons of algorithm performance are given for a variety of models and training dataset sizes, along with the results of experiments demonstrating the ability to produce a custom-defined pulse shape using dispersion-coefficient control, the first step toward arbitrary pulse-shaping control at the femtosecond level for high-power laser applications.
Machine learning is increasingly used in materials discovery, but its practical application is often limited by the time required to construct structured experimental datasets and by the lack of reliable uncertainty estimates. We introduce an integrated workflow that combines document-grounded large language model (LLM) literature mining, human verification, and probabilistic modeling to enable uncertainty-aware prediction of the coefficient of thermal expansion (CTE) in complex oxides. Experimentally reported CTE values, compositions, and measurement temperature ranges are extracted from full-text articles using a document-grounded large language model and subsequently normalized and verified. Composition-derived descriptors and the reported temperature bounds are used as inputs to a probabilistic regression model that predicts both the expected CTE and a composition-dependent uncertainty, enabling prediction intervals for new compositions. On held-out tests, the model achieves competitive mean accuracy relative to deterministic baselines while producing uncertainty estimates that increase systematically for sparsely represented or chemically distinct compositions, enabling risk-aware screening and prioritization. The workflow supports comparative screening of compositions with targeted CTE behavior and helps guide experimental selection for detailed thermophysical characterization. This study illustrates how LLM-assisted literature curation can be combined with uncertainty-aware machine learning to construct property prediction workflows for materials systems where experimental data are sparse and primarily available in the literature.
Charge-transport models provide quantitative and physically interpretable descriptions of memristive devices but are computationally prohibitive for highly iterative tasks in model-driven design workflows such as parameter space exploration, parameter extraction, and optimization. Here, we investigate recurrent neural networks (RNNs) as efficient sequence-to-sequence surrogates to accelerate dynamic memristive transport models and provide systematic guidance on training strategies, architectural choices, and data requirements. In this context, we focus on long short-term memory (LSTM) and gated recurrent unit (GRU) architectures combined with advanced training and normalization strategies. The results demonstrate that layer normalization substantially improves convergence, training stability, and generalization, whereas chrono initialization degrades performance in this setting. The most robust training behavior is obtained by combining layer normalization with the AdamW optimizer and cosine annealing learning-rate scheduling, with GRU architectures achieving the lowest errors overall. Using the optimized configurations, mean normalized errors below 0.1% are achieved for a five-dimensional parameter space. Accurate performance is retained with limited training data, with fewer than 1000 configurations still yielding mean errors around 0.15%. Increasing the input dimensionality leads to a systematic rise in error from ∼0.1% (5D) to 0.66% (9D), while mean errors remain small, well below 1%. These results establish practical design rules for applying RNN-based surrogates in model-driven design and optimization of memristive devices.
Neuromorphic vision systems, inspired by biological vision, offer high energy efficiency and sensor-level data processing. Among these systems, neuromorphic optoelectronic sensors are particularly promising because they exhibit biomimetic responses to incident light. A key feature of such devices is an inherent memory-like response, hereafter referred to as “photonic memory”: sensor conductance depends not only on instantaneous illumination but also on exposure history, enabling temporal integration of information. Here, we study how photonic memory impacts trajectory prediction by comparing a ZnO-based photonic-memory sensor with a conventional image sensor in an experimentally grounded numerical framework. We experimentally characterize the non-volatile conductivity dynamics of the synapse under controlled illumination and darkness and develop a compact mathematical model of its memory behavior. The resulting imaging model is further validated using experimentally reconstructed motion patterns based on measured ZnO-based synaptic-pixel responses. We then evaluate both sensors in a trajectory prediction task across different noise levels and numbers of input frames supplied to an artificial neural network. Our results show that, within this experimentally grounded numerical benchmark, photonic memory provides a clear advantage in the single-frame regime, where the conventional sensor can rely only on spatial priors rather than explicit temporal information. However, as the number of input frames increases, this advantage diminishes and is eventually lost. A conventional image sensor then achieves superior multi-frame performance. These findings suggest that photonic memory alone is insufficient for optimal multi-frame trajectory prediction and must be complemented by additional biologically inspired mechanisms, such as synaptic-like depression implemented at the level of device physics and sensor architecture. More broadly, our work outlines design directions for optoelectronic vision hardware that exploit a richer and more versatile space of device-level temporal responses.
Explainable artificial intelligence methods are increasingly essential for extracting physical insight from high-throughput computational materials data. We apply the Shapley behavioral transformation framework to two MXene datasets to characterize how individual compositions contribute to dataset statistics. By decomposing variance, skewness, kurtosis, and entropy using Shapley values, we create four complementary “behavioral spaces” that reveal compositional patterns invisible in raw feature representations. All behavioral spaces exhibit strong clustering tendency, with the skewness space providing the clearest chemical interpretation. Regional analysis in skewness space identifies six distinct behavioral zones with characteristic property distributions, including a fluorine-terminated MXenes cluster with elevated intercalation voltages and deep valence band positions. Critically, no MXenes appear as an outlier in all four behavioral spaces, demonstrating that each statistic captures genuinely complementary aspects of distributional behavior. The maximum cross-space agreement is 3/4 spaces, achieved by Sc2C in the electrochemistry dataset and by eight compositions in the electronic structure dataset. Among these consistent outliers, Y2CBr2 and Hf3C2(NH)2 exhibit conduction band positions favorable for photocatalytic hydrogen evolution. This model-agnostic framework complements predictive machine learning by answering not “what property will this composition have?” but “how does this composition contribute to the statistical structure of the dataset?” The latter question is directly relevant for identifying synthesis priorities and understanding structure–property relationships in complex materials spaces.
Two-dimensional (2D) particle systems, such as magnetic skyrmions, exhibit topological phase transitions between unique 2D phases. However, a simple and computationally efficient methodology to capture lattice configurational properties and construct an appropriate, easily calculable indicator for phase identification remains elusive. Here, we propose an indicator for topological phase transitions using persistent homology (PH). PH provides a complementary topological description by capturing persistent features derived from the configurational properties of the lattice. The proposed persistent-homology-based indicator, which selectively counts stable features in a persistence diagram, effectively traces the lattice’s ordering changes, as confirmed by comparisons with the conventionally used measure of the ordering (the magnitude of the orientational order parameter ⟨|Ψ6|⟩), typically used to identify lattice phases. We demonstrate the applicability of our indicator to experimental data, showing that it yields results consistent with those of simulations. This experimental validation highlights the robustness of the proposed method for real physical systems beyond idealized simulated systems. While our method is demonstrated in the context of skyrmion lattice systems, the approach is general and can be extended to other two-dimensional systems composed of interacting particles. The proposed topological indicator remains computationally tractable for the system sizes considered here, while preserving the topological invariant-based structural information.
Accurately predicting ionic conductivity is key to accelerating the discovery of solid-state electrolytes; yet, high-throughput experiments are labor-intensive, and first-principles simulations remain too slow. Previously, machine-learning models have been developed that incorporate structural, potential, or compositional information for prediction. Still, their poor prediction accuracy for the distinct ionic conductivities of polymorphic materials with identical compositions but different crystal structures remains problematic. Here, we present a three-dimensional convolutional neural network that learns from both the three-dimensional geometry and the spatial potential landscape of lithium ions. The former is expressed by the lithium-ion distribution, and the latter is encoded from the Bond Valence Sum Energy to predict ionic conductivity. The proposed model, trained on experimental ionic conductivities, achieves prediction performance comparable to that of a state-of-the-art model that uses composition information, and it ranks the mobilities of several materials sharing the same composition but different structures. Our findings highlight the effectiveness of incorporating structural information into conductivity prediction models to overcome the limitations of conventional approaches, potentially improving the screening of solid electrolytes by utilizing the three-dimensional physical descriptor.
Powder diffraction is a primary structural characterization tool in materials science, yet automated phase identification remains a major bottleneck for autonomous discovery. Existing workflows rely heavily on search–match heuristics and manual Rietveld refinement, and broadly usable end-to-end automation is especially limited for neutron powder diffraction, where comparable tools are largely absent. Here we introduce RADAR-PD, a modality-aware machine learning framework for phase identification and quantification across both X-ray and neutron powder diffraction. RADAR-PD couples a mismatch-tolerant neural network operating on coarse momentum-transfer fingerprints with automated lattice nudging and physics-constrained Rietveld verification, enabling dominant-phase hypotheses to be generated from elemental constraints and secondary phases to be isolated recursively. On an experimental RRUFF PXRD benchmark, RADAR-PD outperforms DARA in recovering the reference phase. RADAR-PD further provides robust multiphase analysis on complex time-of-flight and constant-wavelength neutron datasets, addressing an important unmet need in automated neutron diffraction analysis. These results establish RADAR-PD as an auditable, instrument-agnostic framework for autonomous structural discovery.
Recent advances in ab initio molecular dynamics (AIMD) have enabled precise simulations of vibrational dynamics in molecular systems; however, the high computational cost of AIMD limits its application to small-scale systems and short time spans. Machine learning interatomic potentials (MLIPs) offer a promising route to extend these simulations. Training of MLIPs, however, is usually nontrivial, especially when attempting to capture both the structural and vibrational dynamics of molecular systems. In this paper, we introduce a multi-stage workflow that combines AIMD data from diverse simulation packages to simultaneously train multiple deep-learning models such as DeePMD-kit, NequIP, and Allegro. The framework employs a genetic algorithm for hyperparameter optimization and utilizes inelastic neutron scattering spectra as an additional performance metric, ensuring both the static structure and dynamic behavior are accurately reproduced. This integrated approach not only enhances the reliability of MLIPs in capturing complex interatomic interactions but also paves the way for more predictive and efficient materials modeling.
Degradation is a central bottleneck for deploying next-generation energy devices, yet generating large, long-duration aging datasets is costly and slow. This Perspective surveys small-data machine learning (ML) techniques that extract maximal information from limited measurements to accelerate lifetime prediction, interpretation, and optimization. Using three case studies on batteries, fuel cells, and solar cells, we benchmark feature-engineered regression models and show that simple models using carefully chosen physics-based features can accurately forecast even with small datasets. We then demonstrate how interpretable ML links processing and operating parameters to degradation pathways and how physics-informed features improve model robustness. For optimization under data constraints, we compare Bayesian optimization (BO) and reinforcement learning, highlighting BO as a broadly applicable strategy across composition, manufacturing, and device operation optimization. We further describe data fusion and transfer learning strategies that combine multi-fidelity and multi-laboratory datasets and transfer knowledge across chemistries to mitigate data scarcity. Finally, we outline open challenges and research gaps in data, modeling, as well as hardware and software integration, aiming to motivate continued progress toward data-driven solutions for degradation challenges in energy device research and development.
Novel materials for spintronic devices require not only bulk electronic properties but also electronic structures compatible with magnetic/nonmagnetic interfaces. First-principles calculations have revealed the importance of electron band matching (EBM) at the Fermi energy for transport properties such as giant magnetoresistance (GMR). However, direct evaluation of ballistic conductance in GMR devices using first-principles calculations is computationally expensive and unsuitable for high-throughput screening. Here, we propose a machine learning framework to predict ballistic conductance using image representations of the Fermi surface similarity between magnetic and nonmagnetic materials. The images are decomposed into dictionary atoms, and the resulting feature vectors are subjected to regression analysis, with Gaussian process regression achieving the highest accuracy among the various models. Our results demonstrate that the proposed method effectively captures interfacial features essential for electron transport and provides a rapid means to screen material combinations with large EBM. This framework offers a practical step toward realizing material informatics for spintronics, enabling efficient exploration of heterogeneous material combinations for next-generation GMR devices.
Multipactor is a surface-driven electron avalanche phenomenon that degrades the performance and reliability of radio frequency (RF) systems in particle accelerator and vacuum electronics applications. Multipactor behavior in a given device structure is conventionally assessed through susceptibility charts, which provide a parameter-space characterization of the instability. In this work, we assess the capabilities of machine-learning (ML) models to learn and predict such susceptibility charts and analyze the constraints governing their generalization across materials. Using a simulation-derived dataset spanning six distinct secondary-electron-yield material profiles in a canonical two-surface planar geometry, we train supervised regression models and artificial neural networks to predict the time-averaged electron growth rate, δavg, across the relevant parameter space. Model performance is evaluated using metrics that explicitly probe the structure of susceptibility charts, including intersection over union, structural similarity index, and correlation analysis. Tree-based ensemble models outperform neural-network models in reconstructing susceptibility regions and in generalizing across material domains. Principal-component analysis reveals disjoint material feature distributions, indicating that the piecewise mode structure of multipactor susceptibility is difficult to represent with a single global model and that generalization is constrained by data coverage rather than by model complexity. An exhaustive reduced-coverage study further shows that sparse material-space coverage can yield mean performance in the same general range but producing large variability in the susceptibility-region overlap. These results clarify the capabilities of ML-based surrogate models for parameter-space characterization of multipactor discharge. They also provide guidance for their appropriate use in RF system design.
Aluminum nano- and micro-scale particles are attractive fuel for energetic reactions due to their high energy density, natural abundance, environmental stability, and potential for rapid energy release. However, the mechanisms governing the liberation of the metallic Al core from its native alumina shell remain poorly understood. To address this challenge, we developed an automated experimentation pipeline that first utilizes machine learning-based particle localization and characterization using scanning electron microscopy and then reacts Al particles with a separate autonomous optical microscope-based setup that employs focused pulsed laser irradiation for localized heating. With a human out of the experimentation loop, we achieved more than a 100× improvement in experimental throughput compared with manual operation. As a demonstration of autonomy, the system mapped the laser energy threshold required to induce Al particle explosions as a function of particle diameter (0.4–2.5 μm). These findings represent a critical first step toward autonomous exploration of more complex energetic reactions, in which advanced artificial intelligence-based planning dynamically selects experiments to accelerate mechanistic discovery.
Neutron spin echo (NSE) spectroscopy is a dynamical neutron scattering technique in high demand worldwide. Experiments are long, typically ranging from hours to days, such that small improvements in measurement efficiency translate into substantial reductions in experimental costs. Active learning (AL), in which new measurement conditions are chosen based on the experimental data collected so far, has previously been shown to significantly improve data collection rates for one-dimensional neutron reflectometry patterns. These AL algorithms used Bayesian inference to evaluate the uncertainty in a model of the neutron reflectometry pattern, while the acquisition function selected new measurement conditions that were expected to reduce uncertainty in specific model parameters of interest. Similarly, the NSE response function is a highly nonlinear, one-dimensional damped sinusoid, with the added benefit of an analytical form. Here, we compare the performance of AL algorithms, informed by Bayesian inference and driven by different acquisition functions, for measurement of the NSE response function. We find that general acquisition functions designed to most efficiently reduce the global uncertainty in the model function perform well in large search spaces, while AL algorithms based on acquisition functions that highlight specific parameters significantly outperform the general models at that task. For the NSE function, the AL algorithm is robust even against vanishingly small signals. Finally, fast convergence of the global function does not necessarily imply fast convergence of any individual parameter describing it; thus, if possible, it is important to identify and optimize the specific parameters of interest.
Engineering electron correlations in quantum dot arrays demand navigation of high-dimensional, non-convex parameter spaces, where hole doping fundamentally alters the physics. We present a rigorous comparative study of two control paradigms for the 1-hole of half-filled Hubbard model: (i) systematic physics-guided design and (ii) autonomous deep reinforcement learning (RL) with geometry-aware neural architectures. While systematic analysis reveals key design principles—such as field-induced localization for trapping the mobile hole—it is computationally intractable for optimization. We demonstrate that an autonomous RL agent, benchmarked across five 3D lattices (tetrahedron to FCC), achieves human-competitive accuracy (R2 > 0.97) and 95.5% success on held-out tasks. Critically, the RL agent achieves this performance with 103−4× greater sample efficiency than grid search and outperforms other black-box optimization methods. Transfer learning demonstrates 91% few-shot generalization to unseen geometries. This work establishes autonomous RL as a viable, highly efficient framework for rapid optimization and non-obvious strategy discovery in complex quantum systems.
The integration of machine learning into fluid dynamics has accelerated in recent years, driven by the proliferation of high-fidelity data and enhanced computational resources. Acting as efficient surrogate models for computationally intensive simulations, these data-driven approaches provide substantial benefits, particularly during the preliminary stages of design and optimization. Previous investigations have employed convolutional neural networks (CNNs) to predict thermo-fluid flow properties for a variety of channel geometries. These studies have largely relied on data augmentation techniques to handle geometric transformations. However, such augmentation strategies are often inefficient in capturing the inherent flip and shift invariances of flow channel data. In this study, we demonstrate that embedding these invariances directly into the model architecture not only enhances robustness but leads to superior performance while significantly reducing the number of parameters compared to their invariant-unaware counterparts. In particular, we introduce two novel architectures designed to alleviate the sensitivity of CNNs to periodic signal shifts and vertical flips. This approach allows the model to structurally address the geometric symmetries of the flow channel data, offering a more robust alternative to standard data augmentation.
We experimentally demonstrate spin-wave-based reservoir computing (RC) using a device with one-input and nine-output antennas on a yttrium iron garnet film, and we study the relationship between nonlinear phenomena and computational capacities to clarify the mechanism of computing. First, a single series of random binary data is converted into 72 different input voltage signals by varying four encoding parameters: amplitude, time-step length, frequency, and duty ratio. These waveforms are then fed into the reservoir. We propose a “nonlinear transformation index (NT)” to quantitatively evaluate the strength of nonlinear phenomena in reservoir output signals. The values of NT are directly estimated from discrete Fourier transformation spectra of input and output signals to reveal that NT changes with the input parameters. Then, RC is performed for short-term memory tasks and temporal exclusive-OR tasks using a series of random binary data. It is found that there is a strong relationship between NT and the computational capacities and that a trade-off between capacities for linear and nonlinear tasks originates from the strength of linear and nonlinear phenomena. As a result, our NT approach can quantitatively identify the relationship between nonlinear phenomena and computational capacities, unlike the qualitative suggestions in previous studies. Furthermore, with the help of our characterization, enhancement of computional performance can be achieved when some input conditions that have moderate NT values (0.4–0.6) are selectively collected. Our systematic analysis can provide further understanding of physical RC as well as a guideline toward the establishment of a practical computing scheme.
