Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Few Shot Learning”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 55 records · Page 3
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Dynamic in-context learning with conversational models for data extraction and materials property prediction

The advent of natural language processing and large language models (LLMs) has revolutionized the extraction of data from unstructured scholarly papers. However, ensuring data trustworthiness remains a significant challenge. In this paper, we introduce PropertyExtractor, an open-source tool that leverages advanced conversational LLMs such as Google gemini-pro and OpenAI gpt-4, blends zero-shot with few-shot in-context learning, and employs engineered prompts for the dynamic refinement of structured information hierarchies—enabling autonomous, efficient, scalable, and accurate identification, extraction, and verification of material property data. Our tests on material data demonstrate precision and recall that exceed 95% with an error rate of ∼9%, highlighting the effectiveness and versatility of the toolkit. Finally, databases for 2D material thicknesses, a critical parameter for device integration, and energy bandgap values are developed using PropertyExtractor. In particular, for the thickness database, the rapid evolution of the field has outpaced both experimental measurements and computational methods, creating a significant data gap. Our work addresses this gap and showcases the potential of PropertyExtractor as a reliable and efficient tool for the autonomous generation of various material property databases, advancing the field.

Ekuma, Chinedu E. (ORCID:0000000258527556)↗
NASA NTRS1993

Classification of multispectral image data by the Binary Diamond neural network and by nonparametric, pixel-by-pixel methods

The classification of multispectral image data obtained from satellites has become an important tool for generating ground cover maps. This study deals with the application of nonparametric pixel-by-pixel classification methods in the classification of pixels, based on their multispectral data. A new neural network, the Binary Diamond, is introduced, and its performance is compared with a nearest neighbor algorithm and a back-propagation network. The Binary Diamond is a multilayer, feed-forward neural network, which learns from examples in unsupervised, 'one-shot' mode. It recruits its neurons according to the actual training set, as it learns. The comparisons of the algorithms were done by using a realistic data base, consisting of approximately 90,000 Landsat 4 Thematic Mapper pixels. The Binary Diamond and the nearest neighbor performances were close, with some advantages to the Binary Diamond. The performance of the back-propagation network lagged behind. An efficient nearest neighbor algorithm, the binned nearest neighbor, is described. Ways for improving the performances, such as merging categories, and analyzing nonboundary pixels, are addressed and evaluated.

Salu, Yehuda↗
DOE OSTI2025

Knowledge Oriented Graph Unified Transformer (KOGUT) v0.1

KOGUT — Knowledge Oriented Graph Unified Transformer KOGUT implements the Relational Graph Transformer (RelGT) architecture for knowledge graph link prediction in biological domains, with a primary focus on microbial growth media prediction. While the original RelGT (arXiv:2505.10960) targets relational tables, time series, and multi-table databases, KOGUT adapts this architecture for heterogeneous biological knowledge graphs, providing first-in-class AI predictive models for microbial cultivation. Key Adaptations Beyond Original RelGT: - Knowledge Graph Focus: Applied to biological KGs with semantic node types (taxa, chemicals, media, phenotypes, environments) versus generic relational database tables, trained on the KG-Microbe knowledge graph (1.3M entities, 2.9M edges, 24 relation types). - Multimodal Node Encoding: Integrates node labels, categories, descriptions, and synonyms from KG metadata through learned embedding layers—adapting relational column features to graph node attributes with textual semantics. - Extended K-Hop Subgraph Strategy: Optimized neighborhood sampling (3-hop default, configurable up to 200 nodes) tuned for sparse biological networks, building on the original local-global attention framework with biological relation preservation. - Biolink Predicate Preservation: Type-specific transformations for 24 biological edge semantics (occurs_in, consumes, produces, has_phenotype, subclass_of) beyond standard relational foreign keys, enabling multi-relation link prediction. - Inductive Learning Support: Enables zero-shot predictions for novel taxa through feature-based embeddings (temperature, oxygen requirements, gram stain, cell shape), extending the original transductive relational benchmark scope to uncultured microorganisms. CheapSOTA Performance Optimizations (This Distribution): - VQ-EMA Centroid Attention: Vector quantization with exponential moving average for improved global context modeling (+5-10% MRR improvement). - HDF5 Precomputed Data Loading: One-time preprocessing of k-hop subgraphs to eliminate redundant graph traversals (2-5× training speedup). - Distributed Data Parallel Training: Multi-GPU support for scaling to larger knowledge graphs (tested on 4× NVIDIA A100 GPUs at NERSC Perlmutter). - Mixed Precision Training: Automatic mixed precision (AMP) for memory efficiency and faster training. Advantages Over Standard Knowledge Graph Embedding Models: Combines RelGT's proven multi-element tokenization (features, type, hop, structure) with graph-native biological representations, enabling interpretable link prediction across heterogeneous entities that standard embedding models (TransE, RotatE, ComplEx) and table-based transformers cannot directly model. Achieves near-perfect performance on microbial growth media prediction (MRR: 0.9966, Precision@1: 0.9932, Hit@10: 1.0000) while maintaining explainability through attention-based reasoning over biological pathways. Training Data: - KG-Microbe merged knowledge graph: 1,379,337 nodes, 2,960,472 edges - 24 biological relation types including taxonomic hierarchies, metabolic interactions, phenotype associations, and environmental relationships - Primary prediction task: Growth media suitability for microbial taxa (biolink:occurs_in, 50K edges) - Multi-relation capability: Predicts links for any of the 24 relation types, including chemical consumption/production, phenotype associations, and taxonomic classification Citation: Original RelGT Architecture: Dwivedi et al., "Relational Graph Transformer", arXiv:2505.10960, 2025 KOGUT Implementation: Knowledge Oriented Graph Unified Transformer for Microbial Growth Media Prediction Developed at Lawrence Berkeley National Laboratory (LBNL) Trained on NERSC Perlmutter supercomputer

Joachimiak, Marcin [Lawrence Berkeley National Lab↗
DOE OSTI2025

Investigating performance and variability of NIF ICF experiments with deep learning

The parameter space involved in designing an inertial confinement fusion shot at the National Ignition Facility (NIF) is massively multi-dimensional and the cost of a single shot makes a comprehensive set of sensitivity studies in the laboratory impractical. The use of machine learning to overcome these challenges has gained popularity and has had several successful applications by the scientific community. We extend on these efforts by training a neural network (NN) on information about the experimental design, engineering elements, and drive asymmetry to predict with uncertainty the neutron yield of an experiment. We find the measured and model predicted values are in good agreement, with an R 2 value of 0.91 for a randomly selected test dataset. Almost all the predicted 95% credible intervals contain the corresponding measured value for both training and test datasets. We identify correlations picked up by the NN between the shot design, yield, and variability and use them to motivate shot sensitivity studies. The first shot to exceed the Lawson-like ignition criteria (N210808) was conducted at the NIF and subsequent shots studied the design’s robustness. In a follow-up shot to N210808, our model predicts capsule quality to be the main performance degradation mechanism that prevented the shot from repeating previous performance levels. Shot N221204 was the first shot to exceed a target energy gain of 1. Our model predicts increased yield with reduced coast time for a N221204 study and greater variability for designs with lower peak powers at constant yield. The model’s fast prediction speed and uncertainty prediction are useful for identifying interesting design paths that could warrant further investigation with conventional simulations to search for robust high yield designs.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
NASA NTRS1980

A low-cost forward fairing for the Bell Long Ranger Helicopter

A description is presented of work concerned with determining the effects of long-term flight service on advanced composite helicopter airframe components. The helicopter chosen for the program is the Long Ranger Model 206L. The components to be evaluated include the baggage door, litter door, vertical fin, and forward fairing. Only the vertical fin is classified as primary structure. Loss of any of the other components will not compromise safety of the aircraft. Attention is given to the program objectives, the design of the forward fairing, the fabrication procedures, the exterior surface, the cocure procedure, material tests, and initial cost-tracking. The considered program demonstrates the ability to produce an acceptable fairing by the 'one-shot' cocured process and, based on learning curve experience, production costs will be low. The low-temperature 200 F cure does not affect the structural properties to an unacceptable degree. A method for obtaining a smooth, exterior painted surface for Kevlar/epoxy fabric has been developed.

Zinberg, H.↗
DOE OSTI2025

femto-PIXAR: a self-supervised neural network method for reconstructing femtosecond X-ray free electron laser pulses

X-ray Free Electron Lasers (X-FELs) operate in a wide range of lasing configurations for a broad variety of scientific applications at ultrafast time-scales such as structural biology, materials science, and atomic and molecular physics. Shot-by-shot characterization of the X-FEL pulses is crucial for analysis of many experiments as well as tuning the X-FEL performance. However, for the weak pulses found in advanced configurations, e.g. those needed for coherent, two-pulse studies of quantum materials, there is no current method for reliably resolving pulse profiles. Here we show that a physics-based U-net model can reconstruct the individual pulse power profiles for sub-picosecond pulse separation without the need for simulations. Using experimental data from weak X-FEL pulse pairs, we demonstrate we can learn the pulse characteristics on a shot-by-shot basis when conventional methods fail.

43 PARTICLE ACCELERATORS↗
DOE OSTI2025

Exploring Capability of Multimodal Foundation Model for Image-based Fault Detection of Photovoltaic Modules

Multimodal Foundation Model (MFM), like ChatGPT and Gemini, have emerged as powerful tools for their exceptional natural language processing capabilities and their emerging potential in image analysis. This paper investigates the application of MFMs for photovoltaic (PV) fault detection through image analysis, focusing on ChatGPT 4.0 and Gemini 1.5 Pro. Three types of PV images and the corresponding common PV faults are detected: bird droppings using visible images, cell cracks via electroluminescence (EL) images, and hotspots using infrared (IR) images. Among the two models, Gemini 1.5 Pro demonstrated superior performance, achieving near-perfect results with an average F1 score of 0.97, consistently outperforming ChatGPT 4.0 in accuracy and reliability. Unlike traditional machine learning (ML) models, MFMs can operate in a zero shot manner that does not require additional training by the user, and the input images are not limited by size, angle, scope, or PV technology. The strong adaptability and user-friendliness make MFM a promising tool for analyzing PV images and advancing health monitoring for PV modules.

Li, Baojie↗
DOE OSTI2025

One-shot gas detection with transformer paired neural networks in Mako collected longwave infrared hyperspectral imagery

To date, careful data treatment workflows and statistical detectors are used to perform hyperspectral image (HSI) detection of any gas contained in a spectral library, which is often expanded with physics models to incorporate different spectral characteristics. In general, surrounding evidence or known gas-release parameters are used to provide confidence in or confirm detection capability, respectively. This makes quantifying detection performance difficult as it is nearly impossible to develop an absolute ground truth for gas target pixel presence in collected HSI. Consequently, developing and comparing new detection methods, especially machine learning (ML)-based methods, is susceptible to subjectivity in derived detection map quality. Here, in this work, we demonstrate the first use of transformer-based paired neural networks (PNNs) for one-shot gas target detection for multiple gases while providing quantitative classification and detection metrics for their use on labeled data. Terabytes of training data are generated from a database of long-wave infrared HSI obtained from historical Mako sensor campaigns over Los Angeles. By incorporating labels, singular signature representations, and a model development pipeline, we can tune and select PNNs to detect multiple gas targets that are not seen in training on a quantitative basis. We additionally assess our test set detections using interpretability techniques widely employed with ML-based predictors, but less common with detection methods relying on learned latent spaces.

Hyperspectral imaging↗
DOE OSTI2026

Generalist multimodal AI: A review of architectures, challenges and opportunities

Multimodal models are expected to be a critical component to future advances in artificial intelligence. Here, this field is starting to grow rapidly with a surge of new design elements motivated by the success of foundation models in natural language processing (NLP) and vision. It is widely hoped that further extending the foundation models to multiple modalities (e.g., text, image, video, sensor, time series, graph, etc.) will ultimately lead to generalist multimodal models, i.e. one model across different data modalities and tasks. However, there is little research that systematically analyzes recent multimodal models (particularly the ones that work beyond text and vision) with respect to the underling architecture proposed. Therefore, this work provides a fresh perspective on generalist multimodal models (GMMs) via a novel architecture and training configuration specific taxonomy. This includes factors such as Unifiability, Modularity, and Adaptability that are pertinent and essential to the wide adoption and application of GMMs. The review further highlights key challenges and prospects for the field and guide the researchers into the new advancements.

Artificial intelligence (AI)↗
DOE OSTI2025

Massively parallel and universal approximation of nonlinear functions using diffractive processors

Nonlinear computation is essential for a wide range of information processing tasks, yet implementing nonlinear functions using optical systems remains a challenge due to the weak and power-intensive nature of optical nonlinearities. Overcoming this limitation without relying on nonlinear optical materials could unlock unprecedented opportunities for ultrafast and parallel optical computing systems. Here, we demonstrate that large-scale nonlinear computation can be performed using linear optics through optimized diffractive processors composed of passive phase-only surfaces. In this framework, the input variables of nonlinear functions are encoded into the phase of an optical wavefront—e.g., via a spatial light modulator (SLM)—and transformed by an optimized diffractive structure with spatially varying point-spread functions to yield output intensities that approximate a large set of unique nonlinear functions–all in parallel. We provide proof establishing that this architecture serves as a universal function approximator for an arbitrary set of bandlimited nonlinear functions, also covering wavelength-multiplexed nonlinear functions as well as multi-variate and complex-valued functions that are all-optically cascadable. Our analysis also indicates the successful approximation of typical nonlinear activation functions commonly used in neural networks, including the sigmoid, tanh, ReLU (rectified linear unit), and softplus. We numerically demonstrate the parallel computation of one million distinct nonlinear functions, accurately executed at wavelength-scale spatial density at the output of a diffractive optical processor. Furthermore, we experimentally validated this framework using in situ optical learning and approximated 35 unique nonlinear functions in a single shot using a compact setup consisting of an SLM and an image sensor. These results establish diffractive optical processors as a scalable platform for massively parallel universal nonlinear function approximation, paving the way for new capabilities in analog optical computing based on linear materials.

Rahman, Md Sadman Sakib [University of California,↗
DOE OSTI2024

Counter Data Paucity through Adversarial Invariance Encoding: A Case Study on Modeling Battery Thermal Runaway

Lithium-ion batteries, widely used for their durability and high energy storage, face the risk of internal short circuits leading to catastrophic thermal runaway events. These events, triggered by external stimuli like mechanical loads, pose safety concerns in applications such as electric vehicles. Detecting and understanding thermal runaway events is crucial, but physics-driven models struggle to explain the non-linear evolution of battery temperature during these events, considering factors like material composition and state-of-charge. Due to the rarity of these events and the cost of data collection, we propose a deep learning (DL) model to predict battery temperature responses during thermal runaway. The challenge lies in the scarcity of data, making traditional DL models prone to overfitting and learning low-quality representations of the complex process.Our approach introduces a novel few-shot architecture that incorporates an adversarially governed invariant encoding process. This architecture aims to distill "invariant" relationships by addressing distributional shifts in data across various battery properties, facilitating the detection of thermal runaway events. Specifically, our results demonstrate that deep learning models conditioned on these "invariant" representations outperform state-of-the-art baselines, achieving a remarkable 96.8% performance improvement in terms of the popular metric MAPE. This framework presents a promising direction for enhancing battery safety modeling, particularly in the context of rare and complex events like thermal runaway. Our code and code and dataset used for the paper are public1.

Tabassum, Anika [ORNL] (ORCID:0000000254600955)↗
NASA NTRSDate not supplied

The Porous Microstructure Analysis (PuMA) software

The open-source Porous Microstructure Analysis (PuMA) software was implemented to offer an efficient framework for determining material characteristics from 3D microstructures. Its development was inspired by progress in X-ray microtomography, an imaging technology that captures the internal structure of materials in 3D, and even in a 4D temporal context. Over recent years, this method has transformed the domain of materials science due to its capability to non-destructively examine material microstructures while presenting digital data about their geometrical details. It has provided insights into materials relevant to several NASA missions, including heatshields, parachute fabrics, meteorites, and other advanced composites. PuMA, in its current version 3, delivers an array of features, spanning from basic geometric insights of a microstructure to intricate anisotropic thermo-elastic and chemical behavior. Specifically, the software evaluates morphological attributes (specific surface area, volume fractions, mean intercept lengths, orientation) and physical characteristics (conductivity, elasticity, permeability, and tortuosity). Additionally, it can model material degradation processes, such as oxidation and surface chemistry interactions. The software can generate synthetic microstructures, from straightforward geometrical designs to intricate woven and non-woven geometries. Coupling material generation and characterization enable parametric studies and sensitivity analysis to optimize the microstructural performance and inform design decisions and reliability assessment based on uncertainty quantification. A recent addition to PuMA includes the TomoSAM plugin, devised to incorporate the cutting-edge Segment Anything Model (SAM). SAM is a promptable deep learning model that can identify objects and create image masks in a zero-shot manner, based only on a few user clicks. The synergy between these tools aids in the segmentation of complex 3D datasets from tomography or other imaging techniques, which would otherwise require a laborious manual segmentation process.

Tomography↗
NASA NTRSDate not supplied

The Porous Microstructure Analysis (PuMA) software

The open-source Porous Microstructure Analysis (PuMA) software was created to offer an efficient framework for determining material properties from 3D microstructures. Its development was inspired by progress in X-ray microtomography, an imaging technology that captures the internal structure of materials in 3D, and even in a 4D temporal context. Over recent years, this method has transformed the domain of materials science due to its capability to non-destructively examine material microstructures while presenting digital data about their geometrical details. It has provided insights into materials relevant to several NASA missions, including heatshields, parachute fabrics, meteorites, and other advanced composites. PuMA, in its current version 3, delivers an array of features, spanning from basic geometric insights of a microstructure to intricate anisotropic thermo-elastic and chemical behavior. Specifically, the software evaluates morphological attributes (specific surface area, volume fractions, mean intercept lengths, orientation) and physical characteristics (conductivity, elasticity, permeability, and tortuosity). Additionally, it can model material degradation processes, such as oxidation and surface chemistry interactions. The software can generate synthetic microstructures, from straightforward geometrical designs to intricate woven and non-woven geometries. Coupling material generation and characterization enables parametric studies and sensitivity analysis to optimize the microstructural performance and inform design decisions and reliability assessment based on uncertainty quantification. A recent addition to PuMA includes the TomoSAM plugin, devised to incorporate the cutting-edge Segment Anything Model (SAM) into our image segmentation workflow. SAM is a promptable deep learning model that can identify objects and create image masks in a zero-shot manner, based only on a few user clicks. The synergy between these tools aids in the segmentation of complex 3D datasets from tomography and other imaging techniques, which would otherwise require a laborious manual segmentation process.

Tomography↗
DOE OSTI2025

Towards a Robust Adaptive Digital Twin for Fusion Applications

The development of a digital twin system for fusion applications is essential for enhancing the prediction, analysis, and optimization of complex plasma processes. Machine learning (ML), particularly deep learning has demonstrated strong capabilities in modeling such highly nonlinear and intricate systems. However, two critical challenges limit the deployment of deep learning-based digital twins: Uncertainty Quantification (UQ) and data drift. UQ is vital for ensuring trustworthy predictions, especially in decision-support scenarios. Additionally, data-driven models are often sensitive to changes in the underlying data distribution, such as shot-to-shot variations in fusion experiments, which can lead to performance degradation over time. To address these challenges, we are developing an uncertainty-aware, adaptive digital twin framework. Our approach incorporates deep learning models enhanced with Gaussian Process approximations for predictive uncertainty estimation, coupled with an online learning mechanism that enables continuous model adaptation to new experimental data. This adaptive capability allows the data driven models to respond effectively to evolving plasma behaviors and equipment conditions. Specifically, to mitigate the effects of shot-to-shot drift, our system updates itself incrementally as new data becomes available, improving both robustness and fidelity. Our vision is to evolve this data driven model into a self-sustaining digital twin system that leverages UQ based feedback to continuously refine itself and potentially support real-time decision making. This presentation will cover a brief background on uncertainty quantification for ML, our ongoing effort on development of UQ capabilities for ML, our data science pipeline from data collection to model development and analysis and online learning framework for modeling coil deflection at DIII-D. I will also briefly touch upon opportunities and challenges in development of digital twin framework.

Sammuli, Brian [General Atomics]↗
DOE OSTI2026

Robust Design Under Uncertainty in Quantum Error Mitigation

Error mitigation techniques are crucial to achieving near-term quantum advantage. Classical postprocessing of quantum computation outcomes is a popular approach for error mitigation, which includes methods, such as zero noise extrapolation, virtual distillation, and learning-based error mitigation. However, these techniques have limitations due to the propagation of uncertainty resulting from the finite shot number of a quantum measurement. In this work, we introduce general and unbiased methods for quantifying the uncertainty and error of error-mitigated observables based on the strategic sampling of error mitigation outcomes. We then extend our approach to demonstrate the optimization of performance and robustness of error mitigation under uncertainty. To illustrate our methods, we apply them to zero noise extrapolation and Clifford date regression in the ground state of the XY model simulated using depolarizing and International Business Machines Corporation (IBM) Toronto noise models, respectively. In particular, we optimize the choice of noise levels and the allocation of shots for zero noise extrapolation and the distribution of the training circuits for Clifford data regression. While our methods are readily applicable to any postprocessing-based error mitigation approach, in practice they must not be prohibitively expensive—even though they perform optimizations of the error mitigation hyperparameters requiring sampling of a statistical distribution of error mitigation outcomes. By leveraging surrogate-based optimization, we show that our methods can efficiently perform optimal design for a zero noise extrapolation implementation. We then further demonstrate the transferability of learned zero noise extrapolation hyperparameters to other similar circuits.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2024

Linking Threat Agents to Targeted Organizations: A Pipeline for Enhanced Cybersecurity Risk Metrics

In this study, we present a methodology leveraging Large Language Models (LLMs) to transform Cybersecurity Threat Intelligence (CTI) narratives into actionable insights for individual organizations. Our approach automates the extraction of machine-readable adversary SKRAM (Skills, Knowledge, Resources, Authorities, and Motivation) attributes from open-source reports, extending LLM utility beyond typical interactions. This innovation enables precise, automated assessments of cybersecurity risks posed by various adversaries. Using a chain-of-thought and multi-shot prompting strategy, our methodology advances the automation of cybersecurity feature extraction for new machine-learning models that predict the risk of adversary targeting. This approach is refined using a substantial dataset of over 150 analyst-validated threat reports and synthetic organizational data from 900 companies. Here, by bootstrapping the training data with a rule-based heuristic over synthetic data, we have developed a high-accuracy machine-learning model that allows entities to dynamically prioritize threats and defensive actions.

Cyber Threat Intelligence↗
DOE OSTI2025

Temporal Forecasting of Distributed Temperature Sensing in a Thermal Hydraulic System With Machine Learning and Statistical Models

We benchmark performance of long-short term memory (LSTM) network machine learning model and autoregressive integrated moving average (ARIMA) statistical model in temporal forecasting of distributed temperature sensing (DTS). Data in this study consists of fluid temperature transient measured with two co-located Rayleigh scattering fiber optic sensors (FOS) in a forced convection mixing zone of a thermal tee. We treat each gauge of a FOS as an independent temperature sensor. We first study prediction of DTS time series using Vanilla LSTM and ARIMA models trained on prior history of the same FOS that is used for testing. The results yield maximum absolute percentage error (MaxAPE) and root mean squared percentage error (RMSPE) of 1.58% and 0.06% for ARIMA, and 3.14% and 0.44% for LSTM, respectively. Next, we investigate zero-shot forecasting (ZSF) with LSTM and ARIMA trained on history of the co-located FOS only, which is advantageous when limited training data is available. The ZSF MaxAPE and RMSPE values for ARIMA are comparable to those of the Vanilla use case, while the error values for LSTM increase. We show that in ZSF, performance of LSTM network can be improved by training on most correlated gauges between the two FOS, which are identified by calculating the Pearson correlation coefficient. The improved ZSF MaxAPE and RMSPE for LSTM are 4.4% and 0.33%, respectively. Performance of ZSF LSTM can be further enhanced through transfer learning (TL), where LSTM is re-trained on a subset of the FOS that is the target of forecasting. We show that LSTM pre-trained on correlated dataset and re-trained on 30% of testing target dataset achieves MaxAPE and RMSPE values of 2.32% and 0.28%, respectively.

ARIMA↗
DOE OSTI2025

A framework to evaluate machine learning crystal stability predictions

The rapid adoption of machine learning in various scientific domains calls for the development of best practices and community agreed-upon benchmarking tasks and metrics. We present Matbench Discovery as an example evaluation framework for machine learning energy models, here applied as pre-filters to first-principles computed data in a high-throughput search for stable inorganic crystals. We address the disconnect between (1) thermodynamic stability and formation energy and (2) retrospective and prospective benchmarking for materials discovery. Alongside this paper, we publish a Python package to aid with future model submissions and a growing online leaderboard with adaptive user-defined weighting of various performance metrics allowing researchers to prioritize the metrics they value most. To answer the question of which machine learning methodology performs best at materials discovery, our initial release includes random forests, graph neural networks, one-shot predictors, iterative Bayesian optimizers and universal interatomic potentials. We highlight a misalignment between commonly used regression metrics and more task-relevant classification metrics for materials discovery. Accurate regressors are susceptible to unexpectedly high false-positive rates if those accurate predictions lie close to the decision boundary at 0 eV per atom above the convex hull. The benchmark results demonstrate that universal interatomic potentials have advanced sufficiently to effectively and cheaply pre-screen thermodynamic stable hypothetical materials in future expansions of high-throughput materials databases.

Riebesell, Janosh↗