Search NASA⌕ Search

SEARCH · Search NASA

Results for “Adaptive Machine 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 217 records · Page 12
All sourcesNASA NTRSDOE OSTI
DOE OSTI2024

Investigating explainable transfer learning for battery lifetime prediction under state transitions

Battery lifetime prediction at early cycles is crucial for researchers and manufacturers to examine product quality and promote technology development. Machine learning has been widely utilized to construct data-driven solutions for high-accuracy predictions. However, the internal mechanisms of batteries are sensitive to many factors, such as charging/discharging protocols, manufacturing/storage conditions, and usage patterns. These factors will induce state transitions, thereby decreasing the prediction accuracy of data-driven approaches. Transfer learning is a promising technique that overcomes this difficulty and achieves accurate predictions by jointly utilizing information from various sources. Hence, we develop two transfer learning methods, Bayesian Model Fusion and Weighted Orthogonal Matching Pursuit, to strategically combine prior knowledge with limited information from the target dataset to achieve superior prediction performance. From our results, our transfer learning methods reduce root-mean-squared error by 41% through adapting to the target domain. Furthermore, the transfer learning strategies identify the variations of impactful features across different sets of batteries and therefore disentangle the battery degradation mechanisms and the root cause of state transitions from the perspective of data mining. These findings suggest that the transfer learning strategies proposed in our work are capable of acquiring knowledge across multiple data sources for solving specialized issues.

25 ENERGY STORAGE↗
DOE OSTI2024

Exploring electron-beam induced modifications of materials with machine-learning assisted high temporal resolution electron microscopy

Directed atomic fabrication using an aberration-corrected scanning transmission electron microscope (STEM) opens new pathways for atomic engineering of functional materials. In this approach, the electron beam is used to actively alter the atomic structure through electron beam induced irradiation processes. One of the impediments that has limited widespread use thus far has been the ability to understand the fundamental mechanisms of atomic transformation pathways at high spatiotemporal resolution. Here, we develop a workflow for obtaining and analyzing high-speed spiral scan STEM data, up to 100 fps, to track the atomic fabrication process during nanopore milling in monolayer MoS 2 . An automated feedback-controlled electron beam positioning system combined with deep convolution neural network (DCNN) was used to decipher fast but low signal-to-noise datasets and classify time-resolved atom positions and nature of their evolving atomic defect configurations. Through this automated decoding, the initial atomic disordering and reordering processes leading to nanopore formation was able to be studied across various timescales. Using these experimental workflows a greater degree of speed and information can be extracted from small datasets without compromising spatial resolution. This approach can be adapted to other 2D materials systems to gain further insights into the defect formation necessary to inform future automated fabrication techniques utilizing the STEM electron beam.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Machine-Learning-Based Mapping and Modeling of Solar Energy with Ultra-High Spatiotemporal Granularity

Despite the rapid growth of solar energy, we still lack a dynamic, high-fidelity database that tracks the spatiotemporal variations of solar PVs and their associated infrastructures across different places at a spatially resolved scale. The absence of such data presents a barrier to various applications such as solar PV growth projection, solar energy integration, solar incentive design, and climate risk assessment. In this project, we aim to bridge this gap by developing AI-based algorithms to extract granular information about solar PV installations and their associated infrastructures (i.e., distribution grids) from widely available unstructured data like remote sensing images and street views. As a result, we have built the Solar Energy Atlas, a fine-grained, large-scale geospatial overlay of distributed solar PVs and distribution grids. On top of it, we have advanced the understanding of solar adoption and distribution grid vulnerability to climate-induced extremes. Our major contributions can be summarized as follow: (1) By developing new AI algorithms, we have built the most comprehensive solar PV spatiotemporal database covering the entire US. This is the first time we obtained the exact GPS locations, size, subtype, and installation year information for rooftop solar PVs across the US. This database can be used for solar PV growth projection, solar energy integration, solar energy policy analysis and design, and spatially-resolved climate risk assessment. (2) Leveraging this database, we have uncovered the socioeconomic driving factors that are correlated with earlier onset of solar adoption and higher saturated adoption levels. We have identified the heterogeneity in the effects of different types of financial incentives on solar adoption and provided implications for tailoring incentive design based on local income levels to promote equitable solar adoption. (3) We have developed a distribution grid GIS mapping algorithm which can obtain granular geospatial and topology information about distribution grids using multi-modal open data, reducing the dependency on hard-to-obtain smart meter data of conventional approaches. It shows effectiveness in both the U.S. and Sub-Saharan Africa. Using this algorithm, we have uncovered the non-uniform vulnerability of distribution grids to wildfires in California in the aspects of undergrounding protection and Distributed Energy Resources (DER) preparedness. This has provided important implications for improving the affordability and equity of grid adaptation approaches. (3) We have made our produced database publicly available and provided user-friendly interface to enable various stakeholders and the general public to interact with the data. We have also integrated the produced data into the Data Commons platform to enable the public to access the data and correlate it with other location-specific characteristics simply using natural language as queries. The impact of our project is three-fold: (1) New algorithms for mapping solar PVs and distribution grids across space and time, which are open source to facilitate researchers and industry; (2) New databases of solar PVs and distribution grids that have been made publicly available for engineering, social, and policy applications; (3) New understandings and actionable insights on the potential approaches to promoting solar adoption and reducing energy infrastructure vulnerabilities. In this report, we start by discussing the project background and motivation (section 5), followed by the overview of project objectives (section 6). Results and discussion for each task are presented in section 7. Significant accomplishments are summarized in section 8. This report will be concluded by discussing the paths forwards (section 9), products (section 10), and team roles (section 11).

14 SOLAR ENERGY↗
DOE OSTI2024

Unsupervised multimodal fusion of in-process sensor data for advanced manufacturing process monitoring

Effective monitoring of manufacturing processes is crucial for maintaining product quality and operational efficiency. Modern manufacturing environments often generate vast amounts of complementary multimodal data, including visual imagery from various perspectives and resolutions, hyperspectral data, and machine health monitoring information such as actuator positions, accelerometer readings, and temperature measurements. However, fusing and interpreting this complex, high-dimensional data presents significant challenges, particularly when labeled datasets are unavailable or impractical to obtain. This paper presents a novel approach to multimodal sensor data fusion in manufacturing processes, inspired by the Contrastive Language-Image Pre-training (CLIP) model. We leverage contrastive learning techniques to correlate different data modalities without the need for labeled data, overcoming limitations of traditional supervised machine learning methods in manufacturing contexts. Our proposed method demonstrates the ability to handle and learn encoders for five distinct modalities: visual imagery, audio signals, laser position (x and y coordinates), and laser power measurements. By compressing these high-dimensional datasets into low-dimensional representational spaces, our approach facilitates downstream tasks such as process control, anomaly detection, and quality assurance. The unsupervised nature of our method makes it broadly applicable across various manufacturing domains, where large volumes of unlabeled sensor data are common. We evaluate the effectiveness of our approach through a series of experiments, demonstrating its potential to enhance process monitoring capabilities in advanced manufacturing systems. This research contributes to the field of smart manufacturing by providing a flexible, scalable framework for multimodal data fusion that can adapt to diverse manufacturing environments and sensor configurations. The proposed method paves the way for more robust, data-driven decision-making in complex manufacturing processes.

Contrastive Learning↗
DOE OSTI2026

Emerging anomaly detection techniques for electronic health records: A survey

Background Anomaly detection in electronic health records (EHRs) is a cornerstone of biomedical informatics, with direct implications for patient safety, clinical decision-making, and the prevention of healthcare fraud. Once guided primarily by simple rule-based methods, the field has advanced rapidly, driven by increased computing power, richer and more detailed health data, and the rise of machine learning and deep learning techniques. The objective of this paper is to provide a comprehensive overview of modern approaches to detecting anomalies in EHRs, outlining their strengths, limitations, and relevance to key healthcare challenges. We review traditional statistical methods alongside newer ML- and DL-based strategies and hybrid models, with particular attention to how these techniques support transparency and build clinical trust. Methods This paper presents a thorough and critical survey through systematic review (PRISMA-based) of the latest anomaly detection strategies in time-sequence data domains within electronic health record systems. Results We explore a broad spectrum of methodologies, including statistical models, supervised and unsupervised learning approaches, hybrid frameworks, and state-of-the-art ML-based techniques that collectively advance the precision and scalability of detecting anomalies in complex clinical datasets. In addition to mapping current capabilities, we address the enduring challenges that hinder widespread implementation and provide a forward-looking perspective on the future of anomaly detection in the data-rich landscape of modern healthcare. Summary The advancement in AI-based approaches is reported along with the basic principles of the individual approaches and their applicability. The increased availability of high-quality data, advancements in DL approaches, and enhanced computation power are leading to more frequent adaptation of DL-based approaches. Emerging DL-based approaches that have been adapted in other domains or recently applied in the EHR domain are also discussed in detail. Although DL-based approaches can improve model predictions by incorporating comorbidities, their application is limited in low-frequency data domains (e.g., when the total available data remains in the single digits). Therefore, the user must carefully consider the application based on data availability.

Anomaly detection↗
DOE OSTI2026

A Semi-Supervised Learning Method for the Identification of Bad Exposures in Large Imaging Surveys

As the data volume of astronomical imaging surveys rapidly increases, traditional methods for image anomaly detection, such as visual inspection by human experts, are becoming impractical. We introduce a machine-learning-based approach to detect poor-quality exposures in large imaging surveys, with a focus on the DECam Legacy Survey (DECaLS) in regions of low extinction (i.e., E ( B − V ) < 0.04 ). Our semi-supervised pipeline integrates a vision transformer (ViT), trained via self-supervised learning (SSL), with a k-Nearest Neighbor (kNN) classifier. We train and validate our pipeline using a small set of labeled exposures observed by surveys with the Dark Energy Camera (DECam). A clustering-space analysis of where our pipeline places images labeled in good and bad categories suggests that our approach can efficiently and accurately determine the quality of exposures. Applied to new imaging being reduced for DECaLS Data Release 11, our pipeline identifies 780 problematic exposures, which we subsequently verify through visual inspection. Being highly efficient and adaptable, our method offers a scalable solution for quality control in other large imaging surveys.

Luo, Yufeng (ORCID:0000000246230683)↗
DOE OSTI2025

FPGA-accelerated SpeckleNN with SNL for real-time X-ray single-particle imaging

We present the implementation of a specialized version of our previously published unified embedding model, SpeckleNN, for real-time speckle pattern classification in X-ray Single-Particle Imaging (SPI), using the SLAC Neural Network Library (SNL) on an FPGA platform. This hardware realization transitions SpeckleNN from a prototypic model into a practical edge solution, optimized for running inference near the detector in high-throughput X-ray free-electron laser (XFEL) facilities, such as those found at the Linac Coherent Light Source (LCLS). To address the resource constraints inherent in FPGAs, we developed a more specialized version of SpeckleNN. The original model, which was designed for broader classification across multiple biological samples, comprised ~5.6 million parameters. The new implementation, while reducing the parameter count to 64.6K (a 98.8% reduction), focuses on maintaining the model's essential functionality for real-time operation, achieving an accuracy of 90%. Furthermore, we compressed the latent space from 128 to 50 dimensions. This implementation was demonstrated on the KCU1500 FPGA board, utilizing 71% of available DSPs, 75% of LUTs, and 48% of FFs, with an average power consumption of 9.4W according to the Vivado post-implementation report. The FPGA performed inference on a single image with a latency of 45.015 microseconds at a 200 MHz clock rate. In comparison, running the same inference on an NVIDIA A100 GPU resulted in an average power consumption of ~73W and an image processing latency of around 400 microseconds. Our FPGA-accelerated version of SpeckleNN demonstrated significant improvements, achieving an 8.9 × speedup and a 7.8 × reduction in power consumption compared to the GPU implementation. Key advancements include model specialization and dynamic weight loading through SNL, which eliminates the need for time-consuming FPGA design re-synthesis, allowing fast and continuous deployment of models (re)trained online. These innovations enable real-time adaptive classification and efficient vetoing of speckle patterns, making SpeckleNN more suited for deployment in XFEL facilities. This implementation has the potential to significantly accelerate SPI experiments and enhance adaptability to evolving experimental conditions.

47 OTHER INSTRUMENTATION↗
DOE OSTI2026

Generative Models for Crystalline Materials

Understanding structure-property relationships in materials is fundamental in condensed matter physics and materials science. Over the past few years, machine learning (ML) has emerged as a powerful tool for advancing this understanding and accelerating materials discovery. Early ML approaches primarily focused on constructing and screening large material spaces to identify promising candidates for various applications. More recently, research efforts have increasingly shifted toward generating crystal structures using end-to-end generative models. This review analyzes the current state of generative modeling for crystal structure prediction and de novo generation. It examines crystal representations, outlines the generative models used to design crystal structures, and evaluates their respective strengths and limitations. Furthermore, the review highlights experimental considerations for evaluating generated structures and provides recommendations for suitable existing software tools. Emerging topics, such as modeling disorder and defects, integration in advanced characterization, incorporating synthetic feasibility constraints, and model explainability are explored. Ultimately, this work aims to inform both experimental scientists looking to adapt suitable ML models to their specific circumstances and ML specialists seeking to understand the unique challenges related to inverse materials design and discovery.

Metni, Houssam [Karlsruhe Inst. of Technology (KIT↗
DOE OSTI2024

Application of the AI2 Climate Emulator to E3SMv2's Global Atmosphere Model, With a Focus on Precipitation Fidelity

Abstract Can the current successes of global machine learning‐based weather simulators be generalized beyond 2‐week forecasts to stable and accurate multiyear runs? The recently developed AI2 Climate Emulator (ACE) suggests this is feasible, based upon 10‐year simulations with a network trained on output from a physics‐based global atmosphere model using a grid spacing of approximately 110 km and forced by a repeating annual cycle of sea‐surface temperature. Here we show that ACE, without modification, can be trained to emulate another major atmospheric model, EAMv2, run at a comparable grid spacing for at least 10 years with similarly small climate biases—a prerequisite to wider applicability. With an analysis that combines multiple temporal, spatial, and frequency domain perspectives, we show that ACE faithfully represents the spatiotemporal structure of EAMv2 precipitation and related variables. Finally, we show that a pretrained ACE network is able to adapt to a new global climate model simulation data set with 10 fewer training steps than when starting from random initialization, all while still maintaining low levels of climate bias. Further analysis of these fine‐tuning experiments reveal ACE's intriguing ability to interpolate between distinct global climate models.

Duncan, James P. C.↗
DOE OSTI2025

FM4NPP: A Scaling Foundation Model for Nuclear and Particle Physics

Large language models have revolutionized artificial intelligence by enabling large, generalizable models trained through self-supervision. This paradigm has inspired the development of scientific foundation models (FMs). However, applying this capability to experimental particle physics is challenging due to the sparse, spatially distributed nature of detector data, which differs dramatically from natural language. This work addresses if an FM for particle physics can scale and generalize across diverse tasks. We introduce a new dataset with more than 11 million particle collision events and a suite of downstream tasks and labeled data for evaluation. We propose a novel self-supervised training method for detector data and demonstrate its neural scalability with models that feature up to 188 million parameters. With frozen weights and task-specific adapters, this FM consistently outperforms baseline models across all downstream tasks. The performance also exhibits robust data-efficient adaptation. Further analysis reveals that the representations extracted by the FM are task-agnostic but can be specialized via a single linear mapping for different downstream tasks.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2026

Thermodynamics-guided machine learning model for predicting convective boundary layer height and its multi-site applicability

Accurate estimation of convective boundary layer height (CBLH) is vital for weather, climate, and air quality modeling. Machine learning (ML) shows promise in CBLH prediction, but input parameter selection often lacks physical grounding, limiting generalizability. This study introduces a novel ML framework for CBLH prediction, integrating thermodynamic constraints and the diurnal CBLH cycle as an implicit physical guide. Boundary layer growth is modeled as driven by surface heat fluxes and atmospheric heat absorption represented with the low tropospheric stability, using the diurnal cycle as input and output. TPOT and AutoKeras are employed to select optimal models, validated against Doppler lidar-derived CBLH data, achieving an R 2 of 0.84 across untrained years. Comparisons of eddy covariance (ECOR) and energy balance Bowen ratio (EBBR) flux measurements show the same prediction capability. Models trained on the ARM SGP C1 site with ECOR data and tested at E37 and E39 yield R 2 values of 0.79 and 0.81, respectively, demonstrating their adaptability. The ML model trained with all sites' data slightly enhances the performance compared with ML models trained over single-site data. The interquartile range for predicted CBLH is consistently narrower than that for DL-derived CBLH, reflecting lower variability in predicted CBLH compared to DL-derived CBLH, which is influenced by additional factors, which are not well represented with the model inputs. The model's generalizability across multiple sites at the ARM SGP site demonstrates its potential for transfer to greater distances, offering a scalable approach for enhancing boundary layer parameterization in atmospheric models.

Chu, Yufei [Stony Brook Univ., NY (United States)]↗
DOE OSTI2024

Effects of input gradient regularization on neural networks time-series forecasting of thermal power systems

This study proposes using neural networks, specifically gated recurrent unit (GRU), long-short-term memory (LSTM), and transformer networks, to improve control strategies in a 450 MW coal-fired power plant. However, neural networks face issues of becoming overly dependent on just a few variables to make predictions, which negatively impacts control decisions that rely on the model to determine the value of all manipulated variables. The paper introduces regularization techniques, including noise injection and input gradient regularization, during the training phase. Here, the work presents novel contributions in adapting neural networks to control industrial systems and applying regularization techniques from computer vision to industrial process control. Results demonstrate the effectiveness of input gradient regularization in reducing model dependence on subsets of variables, emphasizing the balance between fidelity and controllability. Further exploration is recommended, including the development of recurrent transformers, closed-loop control testing, and a sensitivity analysis on computer models to provide further insight.

20 FOSSIL-FUELED POWER PLANTS↗
DOE OSTI2025

Accelerating computational fluid dynamics simulation of post-combustion carbon capture modeling with MeshGraphNets

Packed columns are commonly used in post-combustion processes to capture CO 2 emissions by providing enhanced contact area between a CO 2 -laden gas and CO 2 -absorbing solvent. To study and optimize solvent-based post-combustion carbon capture systems (CCSs), computational fluid dynamics (CFD) can be used to model the liquid–gas countercurrent flow hydrodynamics in these columns and derive key determinants of CO 2 -capture efficiency. However, the large design space of these systems hinders the application of CFD for design optimization due to its high computational cost. In contrast, data-driven modeling approaches can produce fast surrogates to study large-scale physics problems. We build our surrogates using MeshGraphNets (MGN), a graph neural network framework that efficiently learns and produces mesh-based simulations. We apply MGN to a random packed column modeled with over 160K graph nodes and a design space consisting of three key input parameters: solvent surface tension, inlet velocity, and contact angle. Our models can adapt to a wide range of these parameters and accurately predict the complex interactions within the system at rates over 1700 times faster than CFD, affirming its practicality in downstream design optimization tasks. This underscores the robustness and versatility of MGN in modeling complex fluid dynamics for large-scale CCS analyses.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

An adaptive, data-driven multiscale approach for dense granular flows

The accuracy of coarse-grained continuum models of dense granular flows is limited by the lack of high-fidelity closure models for granular rheology. One approach to addressing this issue, referred to as the hierarchical multiscale method, is to use a high-fidelity fine-grained model to compute the closure terms needed by the coarse-grained model. The difficulty with this approach is that the overall model can become computationally intractable due to the high computational cost of the high-fidelity model. In this work, we describe a multiscale modeling approach for dense granular flows that utilizes neural networks trained using high-fidelity discrete element method (DEM) simulations to approximate the constitutive granular rheology for a continuum incompressible flow model. Our approach leverages an ensemble of neural networks to estimate predictive uncertainty that allows us to determine whether the rheology at a given point is accurately represented by the neural network model. Additional DEM simulations are only performed when needed, minimizing the number of additional DEM simulations required when updating the rheology. This adaptive coupling significantly reduces the overall computational cost of the approach while controlling the error. In addition, the neural networks are customized to learn regularized rheological behavior to ensure well-posedness of the continuum solution. We first validate the approach using two-dimensional steady-state and decelerating inclined flows. We then demonstrate the efficiency of our approach by modeling three-dimensional sub-aerial granular column collapse for varying initial column aspect ratios, where our multiscale method compares well with the computationally expensive computational fluid dynamics (CFD)-DEM simulation.

Dense granular flows↗
DOE OSTI2026

Machine learning at the Spallation Neutron Source accelerator and target

We describe the ongoing efforts to apply Machine Learning techniques to improve the performance of our accelerator and target. Specially, we are looking to minimize halo beam losses in the absence of a proper physics model, automatically detect and log anomalies in the target support systems such as cooling, and detect and prevent errant beam pulses in the linac. We also describe the infrastructure we use to acquire and stream data to the GPU cluster for training, our code development cycle, and edge computing for model inference. To minimize halo beam losses, we use a Reinforcement Learning technique tested on a virtual accelerator. The target anomaly detection is trained on archived data using incomplete physics models and is made part of the existing target reporting system. The errant beam prevention analyzes beam current and beam phase waveforms as well as accelerator configuration data to predict errant pulses. We also develop continual learning to adapt to changes in the accelerator.

Accelerator Physics↗
DOE OSTI2024

Demonstration and Evaluation of Explainable and Trustworthy Predictive Technology for Condition-based Maintenance

The domestic nuclear power plant (NPP) fleet has historically relied on labor-intensive and time-consuming predictive maintenance (PdM) programs, thus driving up operation and maintenance (O&M) costs to achieve high-capacity factors. Artificial intelligence (AI) and machine-learning (ML) can help simplify complex problems such as diagnosing equipment degradation to enable more effective decision-making efforts. The benefits of AI will be felt through more efficient plant O&M, improved work processes, and better integration of people and technology. Together, these benefits hold the promise to make nuclear power more sustainable by reducing O&M costs while improving employee engagement. While AI and ML technologies hold significant promise for the nuclear industry, there are challenges or barriers to their adoption. Explainability and trustworthiness of AI are two salient challenges that need to be addressed for wider deployment of these technologies in NPPs. This research focuses specifically on addressing the explainability and trustworthiness of AI technologies to advance the human, technical, and organization (HTO) readiness levels in adopting a risk-informed PdM strategy at commercial NPPs. In addition, this approach can be adapted to enhance the acceptability of AI in other nuclear applications with a few application-specific modifications. The technical approach ensuring wider adoption of AI technologies was developed by Idaho National Laboratory (INL)—in collaboration with Public Service Enterprise Group (PSEG), Nuclear, LLC—by utilizing the circulating water system (CWS) at two PSEG-owned plant sites for demonstration. Focused user studies were performed in collaboration with subject matter experts (SMEs) from PSEG and other nuclear domains to enhance human and organization readiness by building trust in AI-informed technologies. VIsualization for PrEdictive maintenance Recommendation (VIPER)—a Battelle Energy Alliance, LLC, copyrighted software—was developed and expanded to provide a user-centric visualization by incorporating inputs from the collaborating utility, human factors engineering guidelines, and data analysts. The VIPER software enables users, who may be unfamiliar with ML in general, to be interactively engaged by asking technical questions about PdM, work orders, diagnosis results and their confidence levels, the kind of data being used, and the types of ML algorithms employed. This interactive engagement enhances explainability and builds trust. One of the enabling accomplishments was the integration of large language models (LLMs), both text-based and vision-based, in the VIPER software.

22 GENERAL STUDIES OF NUCLEAR REACTORS↗
DOE OSTI2025

Active oversight and quality control in standard Bayesian optimization for autonomous experiments

The fusion of experimental automation and machine learning has catalyzed a new era in materials research, prominently featuring Gaussian Process (GP) Bayesian Optimization (BO) driven autonomous experiments. Here we introduce a Dual-GP approach that enhances traditional GPBO by adding a secondary surrogate model to dynamically constrain the experimental space based on real-time assessments of the raw experimental data. This Dual-GP approach enhances the optimization efficiency of traditional GPBO by isolating more promising space for BO sampling and more valuable experimental data for primary GP training. We also incorporate a flexible, human-in-the-loop intervention method in the Dual-GP workflow to adjust for unanticipated results. We demonstrate the effectiveness of the Dual-GP model with synthetic model data and implement this approach in autonomous pulsed laser deposition experimental data. This Dual-GP approach has broad applicability in diverse GPBO-driven experimental settings, providing a more adaptable and precise framework for refining autonomous experimentation for more efficient optimization.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Loss Landscape Analysis for Reliable Quantized ML Models for Scientific Sensing

In this paper, we propose a method to perform empirical analysis of the loss landscape of machine learning (ML) models. The method is applied to two ML models for scientific sensing, which necessitates quantization to be deployed and are subject to noise and perturbations due to experimental conditions. Our method allows assessing the robustness of ML models to such effects as a function of quantization precision and under different regularization techniques -- two crucial concerns that remained underexplored so far. By investigating the interplay between performance, efficiency, and robustness by means of loss landscape analysis, we both established a strong correlation between gently-shaped landscapes and robustness to input and weight perturbations and observed other intriguing and non-obvious phenomena. Our method allows a systematic exploration of such trade-offs a priori, i.e., without training and testing multiple models, leading to more efficient development workflows. This work also highlights the importance of incorporating robustness into the Pareto optimization of ML models, enabling more reliable and adaptive scientific sensing systems.

Baldi, Tommaso [Pisa, Scuola Normale Superiore]↗