Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “hyperparameter”

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 91 records · Page 5
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Analyzing and Exploring Training Recipes for Large-Scale Transformer-Based Weather Prediction

Abstract The rapid rise of deep learning (DL) in numerical weather prediction (NWP) has led to a proliferation of models which forecast atmospheric variables with comparable or superior skill than traditional physics-based NWP. However, among these leading DL models, there is a wide variance in both the training settings and architecture used. Further, the lack of thorough ablation studies makes it hard to discern which components are most critical to success. In this work, we show that it is possible to attain high forecast skill even with relatively off-the-shelf architectures, simple training procedures, and moderate compute budgets. Specifically, we train a minimally modified Swin Transformer V2 (SwinV2) on ERA5 data and find that it attains superior skill in terms of mean-square errors of deterministic forecasts when compared against the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS). Almost all DL–NWP systems share a core set of hyperparameters and design decisions. To aid and expedite future DL–NWP research, we present an in-depth, systematic exploration of different loss functions, model sizes and depths, patch sizes, and multistep training objectives. We also examine the model performance with metrics beyond the typical accuracy (ACC) and RMSE and investigate how the performance scales with model size. Through our open-source code, scoring pipelines, and models, we share our findings on key aspects of the training pipeline. These ablations reduce the necessity for expensive hyperparameter tuning and lower the barrier to entry for future DL–NWP research. Significance Statement This study investigates the potential of using large-scale transformer-based models for weather prediction, showing that it is possible to achieve high forecast accuracy with simpler, off-the-shelf architectures. By training a minimally modified SwinV2 transformer on ERA5 data, we show that the model achieves competitive forecast skill in terms of mean-square error for key variables, outperforming the European Centre for Medium-Range Weather Forecasts’ Integrated Forecasting System (IFS) at all lead times. Our findings suggest that effective training strategies, such as multistep fine-tuning and channel-weighted losses, significantly enhance the model’s performance. However, we also highlight that these improvements come with trade-offs in other areas, such as ensemble spread and high-frequency spatial detail. This work highlights the promise of deep learning in improving weather forecasts, which could lead to better preparedness and response to weather events, ultimately benefiting society by providing more reliable weather predictions.

Willard, Jared D. [Lawrence Berkeley National Labo↗
DOE OSTI2026

HydraGNN_Predictive_GFM_2026 - Ensemble of predictive graph foundation models for atomistic materials modeling

This release contains data and parameters of HydraGNN-based graph foundation models trained as a result of the work published in the pre-print "Exascale Multi-Task Graph Foundation Models for Imbalanced, Multi-Fidelity Atomistic Data" by M. Lupo Pasini et al. (https://arxiv.org/abs/2604.15380). We jointly train on 16 open first-principles datasets (544+ million structures covering 85+ elements) using a multi-task architecture with per-dataset heads and a scalable ADIOS2/DDStore data pipeline. On Frontier, we execute six large-scale DeepHyper hyperparameter optimization campaigns in FP64 and promote the top-performing message-passing models to sustained 2,048-node training, yielding a PaiNN-based lead model. The version of HydraGNN used to generate the outputs provided in this release is HydraGNN v5.0 (https://github.com/ORNL/HydraGNN/releases/tag/v5.0) The list of datasets used for the training of the graph foundation model is the following: 1) Alexandria [1] 2) ANI1x [2] 3) MPTrj [3] 4) Open Catalyst 2020 (OC20) [4] 5) Open Catalyst 2022 (OC22) [5] 6) Open Catalyst 2025 (OC25) [6] 7) Open Direct ir Capture 2023 (ODAC23) [7] 8) Open Materials 2024 (OMat24) [8] 9) Open Molecules 2025 (OMol25) [9] 10) OMol25-neutral (subset of OMol25 that contains only molecules with zero total charge) 11) OMol25-non-neutral (subset of OMol25 that contains only molecules with non-zero total charge) 12) Open Polymers 2026 (OPoly2026) [10] 13) Nabla2DFT [11] 14) QCML [12] 15) QM7X [reference 13] 16) transition1x [14] Dataset references: [1] J. Schmidt et al., “A dataset of 175k stable and metastable materials calculated with the PBEsol and SCAN functionals,” Scientific Data, vol. 9, p. 64, 2022. [2] J. S. Smith et al., “The ANI-1ccx and ANI-1x data sets, coupled-cluster and density functional theory properties for molecules,” Scientific Data, vol. 7, p. 134, 2020. [Online]. Available: https: //www.nature.com/articles/s41597-020-0473-z [3] A. Jain et al., “Commentary: The Materials Project: A materials genome approach to accelerating materials innovation,” APL Materials, vol. 1, no. 1, p. 011002, 07 2013. [Online]. Available: https://doi.org/10.1063/1.4812323 [4] L. Chanussot et al., “Open catalyst 2020 (oc20) dataset and community challenges,” ACS Catalysis, vol. 11, no. 10, pp. 6059–6072, 2021. [Online]. Available: https://doi.org/10.1021/acscatal.0c04525 [5] K. Tran et al., “Open catalyst 2022 (oc22) dataset and challenges for oxidation electrocatalysts,” ACS Catalysis, vol. 13, no. 5, pp. 3066–3084, 2023. [Online]. Available: https://doi.org/10.1021/acscatal.2c05426 [6] S. J. Sahoo et al., “The open catalyst 2025 (oc25) dataset and models for solid-liquid interfaces,” arXiv preprint arXiv:2509.17862, 2025. [Online]. Available: https://arxiv.org/abs/2509.17862 [7] A. Sriram et al., “The open DAC 2023 dataset and challenges for sorbent discovery in direct air capture,” ACS Central Science, vol. 10, no. 5, pp. 923–941, 2024. [8] L. Barroso-Luque et al., “Open materials 2024 (omat24) inorganic materials dataset and models,” 2024. [Online]. Available: https://arxiv.org/abs/2410.12771 [9] D. S. Levine et al., “The open molecules 2025 (OMol25) dataset, evaluations, and models,” 2025. [Online]. Available: https://arxiv.org/abs/2505.08762 [10] D. S. Levine et al., The open polymers 2026 (OPoly26) dataset and evaluations,” arXiv preprint arXiv:2512.23117, 2025. [Online]. Available: https://arxiv.org/abs/2512.23117 [11] K. Khrabrov et al., “Nabla2dft: A universal quantum chemistry dataset of drug-like molecules and a benchmark for neural network potentials,” in NeurIPS 2024 Datasets and Benchmarks Track, 2024. [Online]. Available: https://openreview.net/forum?id=ElUrNM9U8c [12] S. Ganscha et al., “The QCML dataset, quantum chemistry reference data from 33.5M DFT and 14.7B semi-empirical calculations,” Scientific Data, vol. 12, p. 406, 2025. [13] J. Hoja et al., “QM7-X, a comprehensive dataset of quantum-mechanical properties spanning the chemical space of small organic molecules,” Scientific Data, vol. 8, p. 43, 2021. [Online]. Available: https://www.nature.com/articles/s41597-021-00812-2 [14] M. Schreiner et al., “Transition1x - a dataset for building generalizable reactive machine learning potentials,” Scientific Data, vol. 9, p. 779, 2022. The folder "datasets_ADIOS2_format" contains the set of pre-processed datasets in Adaptable I/O System (ADIOS) format (https://www.exascaleproject.org/research-project/adios/) that have been used for the development and training of GFMs in this work. The "datasets_ADIOS2_format" directory contains 2 sub-directories, one for the version "v1" of the datasets and one for the version "v2" of the datasets. The version "v1" of the datasets provides values of the total energy as they are extracted from the original data as it was released by the respective institutions. The version "v2" of the datasets provides values of the energy that have been realigned. The realignment was performed by training a linear regression model that predicts the total energy as a function of the chemical composition of the atomistic structure, and then subtract such prediction from the original value of the total energy. Both folders "v1" and "v2" contain 16 sub-directories, each corresponding to an ADIOS2-formatted dataset The folder "DeepHyper-results" contains the configurational files and model's parameters for all the 186 HPO trials that were successfully completed by the scalable hyperparameter optimization (HPO) runs on Frontier. The content of the folder "DeepHyper-results" I structured as follows: 1) task-list.txt: list of mpnn name, jobid, and deephyper task id 2) gfm_${MPNN}_${JOBID}_0.${TASKID}: run directory with checkpoint files 3) gfm_${MPNN}: deephyper summary directory (*.csv) for each specific MPNN type 4) deephyper-experiment-${JOBID}: output and error logs for each job The file "deephyper-sorted.csv" contains the details of each HydraGNN model built and tested by HPO, obtained by merging the (*.csv) filed from each HPO run executed. Out of all the HPO trials, we selected 10 to continue the training of the respective HydraGNN models. Due to limited computational budget available in the LRN070 allocation we could not complete the training till convergence for all these 10 selected models. The folder "models" contains multiple sub-folders, one per each HydraGNN model trained. Each model sub-folder contains the parameters of each HydraGNN model, with multiple checkpoint-restarts. The list of sub-folders are as follows: 1) multidataset_hpo-BEST1-fp64 2) multidataset_hpo-BEST2-fp64 3) multidataset_hpo-BEST3-fp64 4) multidataset_hpo-BEST4-fp64 5) multidataset_hpo-BEST5-fp64 6) multidataset_hpo-BEST6-fp64 7) multidataset_hpo-BEST7-fp64 8) multidataset_hpo-BEST8-fp64 9) multidataset_hpo-BEST9-fp64 10) multidataset_hpo-BEST10-fp64 Within each one of these folders, additional auxiliary log files are provided with descriptions about how the training proceeded. The lead PaiNN-model is contained inside "multidataset_hpo-BEST6-fp64". The file "mlp_branch_weights" contains the parameters of the multi-layer perceptron (MLP) used to reconcile the predictions of the 16 output decoding heads of the HydragNN architectures. The MLP takes in input the chemical composition of the atomistic structure and predicts averaging weights to linearly mix the predictions of each output decoding head toward consolidating them into a single one. The folder "1.1billion-structure-inference" contains 1.1 billion atomistic structures randomly generated. Each structures is associated with energy and forces predicted with the lead-PaiNN model combined with the MLP model for reconciliation of the multi-branch predictions generated by the 16 output decoding heads. The folder "1.1billion-structure-inference" contains 9,300 (*.tar.gz) subdirectories, one per Frontier compute node used to execute the inference at exascale. Once uncompressed, each (*.tar.gz) subdirectory contains an ADIOS2 (*.bp) file container, where each atomistic structure is stored as a PyTorch-Geometric Data object. The file "export_dataset_environment_variables.sh" contains the environment variables that need to be set before running the HydraGNN code to reproduce the results provided in this dataset release. The code that can be used to load the ADIOS2 files, load HydraGNN models, and run inference is available at: https://github.com/ORNL/HydraGNN/releases/tag/v5.0

36 MATERIALS SCIENCE↗
DOE OSTI2026

HydraGNN_OPF_GFM_2026 - Ensemble of predictive graph foundation models for power grid applications

This dataset supports research on graph foundation models for optimal power flow (OPF) on electric grids using HydraGNN. It contains heterogeneous graph representations of PGLib-OPF cases spanning systems from 14 to 13,659 buses, together with packed HDF5 datasets for pretraining, feasibility classification, and N-1 contingency analysis. The release includes OPF solution data, downstream fine-tuning datasets, pretrained HeteroSAGE and HeteroHEAT model checkpoints, hyperparameter-optimization summaries across multiple heterogeneous GNN architectures, and aggregated fine-tuning results for sample-efficiency studies. The dataset is designed to enable scalable training, evaluation, and transfer-learning studies for OPF surrogate modeling, including node-level AC-OPF solution prediction, graph-level prediction, feasibility classification, operating-condition generalization, and contingency-response tasks.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2022

Finding Electronic Structure Machine Learning Surrogates without Training

A myriad of phenomena in materials science and chemistry rely on quantum-level simulations of the electronic structure in matter. While moving to larger length and time scales has been a pressing issue for decades, such large-scale electronic structure calculations are still challenging despite modern software approaches and advances in high-performance computing. The silver lining in this regard is the use of machine learning to accelerate electronic structure calculations – this line of research has recently gained growing attention. The grand challenge therein is finding a suitable machine-learning model during a process called hyperparameter optimization. This, however, causes a massive computational overhead in addition to that of data generation. We accelerate the construction of machine-learning surrogate models by roughly two orders of magnitude by circumventing excessive training during the hyperparameter optimization phase. We demonstrate our workflow for Kohn-Sham density functional theory, the most popular computational method in materials science and chemistry.

36 MATERIALS SCIENCE↗
DOE OSTI2024

Augmented Human Analysis (AHA)

Radio frequency (RF) signal monitoring generally emphasizes intentionally generated signals, such as WiFi, Bluetooth, or cellular transmissions. However, electronic devices also produce unintended radiated emissions (UREs), which could also be useful in RF spectrum analysis. In either case, deriving intelligence from RF signals is typically a human-intensive process requiring significant domain knowledge. In the Augmented Human Analysis (AHA) project, we investigate the utility of dimensionally aligned signal projection (DASP) and machine learning (ML) algorithms for accelerating RF analysis workflows. We find that while DASP algorithms can indeed highlight signal characteristics relevant for classification tasks, the choice of algorithmic hyperparameters greatly affects performance. To address this challenge, we evaluate the quality of DASP outputs using the silhouette score, which measures how well data points cluster; high silhouette scores indicate good clustering, and thus good hyperparameter values. This approach is critical for machine learning pipelines as the DASP parameters cannot be directly optimized during model training. By identifying good DASP parameters, and thus good DASP outputs, as a preprocessing step, we can decrease the amount of effort required for downstream ML model training. We demonstrate our workflow using a dataset of UREs from common household devices, showing that even without the aid of ML, proper selection of DASP parameters enables clustering by device type.

42 ENGINEERING↗
DOE OSTI2025

SIDDA: SInkhorn Dynamic Domain Adaptation

Modern neural networks (NNs) often do not generalize well in the presence of a "covariate shift"; that is, in situations where the training and test data distributions differ, but the conditional distribution of classification labels remains unchanged. In such cases, NN generalization can be reduced to a problem of learning more domain-invariant features. Domain adaptation (DA) methods include a range of techniques aimed at achieving this; however, these methods have struggled with the need for extensive hyperparameter tuning, which then incurs significant computational costs. In this work, we introduce SIDDA, an out-of-the-box DA training algorithm built upon the Sinkhorn divergence, that can achieve effective domain alignment with minimal hyperparameter tuning and computational overhead. We demonstrate the efficacy of our method on multiple simulated and real datasets of varying complexity, including simple shapes, handwritten digits, and real astronomical observations. SIDDA is compatible with a variety of NN architectures, and it works particularly well in improving classification accuracy and model calibration when paired with equivariant neural networks (ENNs). We find that SIDDA enhances the generalization capabilities of NNs, achieving up to a ≈40% improvement in classification accuracy on unlabeled target data. We also study the efficacy of DA on ENNs with respect to the varying group orders of the dihedral group DN, and find that the model performance improves as the degree of equivariance increases. Finally, we find that SIDDA enhances model calibration on both source and target data--achieving over an order of magnitude improvement in the ECE and Brier score. SIDDA's versatility, combined with its automated approach to domain alignment, has the potential to advance multi-dataset studies by enabling the development of highly generalizable models.

Pandya, Sneh [Northeastern U.]↗
DOE OSTI2024

Optimizing Distributed Training on Frontier for Large Language Models

Large language models (LLMs) have demonstrated remarkable success as foundational models, benefiting various downstream applications through fine-tuning. Loss scaling studies have demonstrated the superior performance of larger LLMs compared to their smaller counterparts. Nevertheless, training LLMs with billions of parameters poses significant challenges and requires considerable computational resources. For example, training a one trillion parameter GPT-style model on 20 trillion tokens requires a staggering 120 million exaflops. This research explores efficient distributed training strategies to extract this computation from Frontier, the world's first exascale supercomputer. We enable and investigate various model and data parallel training techniques, such as tensor parallelism, pipeline parallelism, and sharded data parallelism, to facilitate training a trillion-parameter model on Frontier. We empirically assess these techniques and their associated parameters to determine their impact on memory footprint, communication latency, and GPU's computational efficiency. We analyze the complex interplay among these techniques and find a strategy to combine them to achieve high throughput through hyperparameter tuning. We have identified efficient strategies for training large LLMs of varying sizes through empirical analysis and hyperparameter tuning. For 22 Billion, 175 Billion, and 1 Trillion parameters, we achieved GPU throughputs of 38.38%, 36.14%, and 31.96%, respectively. For the training of the 175 Billion parameter model and the 1 Trillion parameter model, we achieved 100% weak scaling efficiency on 1024 and 3072 Mi250X GPUs, respectively. We also achieved strong scaling efficiencies of 89% and 87% for these two models. We trained these models only tens of iterations instead of training till completion.

Yin, Junqi↗
DOE OSTI2022

BUTTER - Empirical Deep Learning Dataset

The BUTTER Empirical Deep Learning Dataset represents an empirical study of the deep learning phenomena on dense fully connected networks, scanning across thirteen datasets, eight network shapes, fourteen depths, twenty-three network sizes (number of trainable parameters), four learning rates, six minibatch sizes, four levels of label noise, and fourteen levels of L1 and L2 regularization each. Multiple repetitions (typically 30, sometimes 10) of each combination of hyperparameters were preformed, and statistics including training and test loss (using a 80% / 20% shuffled train-test split) are recorded at the end of each training epoch. In total, this dataset covers 178 thousand distinct hyperparameter settings ("experiments"), 3.55 million individual training runs (an average of 20 repetitions of each experiments), and a total of 13.3 billion training epochs (three thousand epochs were covered by most runs). Accumulating this dataset consumed 5,448.4 CPU core-years, 17.8 GPU-years, and 111.2 node-years.

Array↗
DOE OSTI2021

CyBERT: Cybersecurity Claim Classification by Fine-Tuning the BERT Language Model

We introduce CyBERT, a cybersecurity feature claims classifier based on bidirectional encoder representations from transformers and a key component in our semi-automated cybersecurity vetting for industrial control systems (ICS). To train CyBERT, we created a corpus of labeled sequences from ICS device documentation collected across a wide range of vendors and devices. This corpus provides the foundation for fine-tuning BERT’s language model, including a prediction-guided relabeling process. We propose an approach to obtain optimal hyperparameters, including the learning rate, the number of dense layers, and their configuration, to increase the accuracy of our classifier. Fine-tuning all hyperparameters of the resulting model led to an increase in classification accuracy from 76% obtained with BertForSequenceClassification’s original architecture to 94.4% obtained with CyBERT. Furthermore, we evaluated CyBERT for the impact of randomness in the initialization, training, and data-sampling phases. CyBERT demonstrated a standard deviation of ±0.6% during validation across 100 random seed values. Finally, we also compared the performance of CyBERT to other well-established language models including GPT2, ULMFiT, and ELMo, as well as neural network models such as CNN, LSTM, and BiLSTM. The results showed that CyBERT outperforms these models on the validation accuracy and the F1 score, validating CyBERT’s robustness and accuracy as a cybersecurity feature claims classifier.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2022

An Accuracy-Maximization Approach for Claims Classifiers in Document Content Analytics for Cybersecurity

This paper presents our research approach and findings towards maximizing the accuracy of our classifier of feature claims for cybersecurity literature analytics, and introduces the resulting model ClaimsBERT. Its architecture, after extensive evaluations of different approaches, introduces a feature map concatenated with a Bidirectional Encoder Representation from Transformers (BERT) model. We discuss deployment of this new concept and the research insights that resulted in the selection of Convolution Neural Networks for its feature mapping aspects. We also present our results showing ClaimsBERT to outperform all other evaluated approaches. This new claims classifier represents an essential processing stage within our vetting framework aiming to improve the cybersecurity of industrial control systems (ICS). Furthermore, in order to maximize the accuracy of our new ClaimsBERT classifier, we propose an approach for optimal architecture selection and determination of optimized hyperparameters, in particular the best learning rate, number of convolutions, filter sizes, activation function, the number of dense layers, as well as the number of neurons and the drop-out rate for each layer. Fine-tuning these hyperparameters within our model led to an increase in classification accuracy from 76% obtained with BertForSequenceClassification’s original model to a 97% accuracy obtained with ClaimsBERT.

Ameri, Kimia (ORCID:0000000328791871)↗
DOE OSTI2020

Photometry as a Proxy for Stellar Activity in Radial Velocity Analyses

Stellar activity remains a limiting factor in measuring precise planet parameters from radial velocity spectroscopy, not least in the search for Earth-mass planets orbiting in the habitable zones of Sun-like stars. One approach to mitigate stellar activity is to use combined analyses of both radial velocity and time-series photometry. We present an analysis of simultaneous disk-integrated photometry and radial velocity data of the Sun in order to determine the useful limits of a combined analysis. We find that simple periodogram or autocorrelation analysis of solar photometry give the correct rotation period <50% of the time. We therefore use a Gaussian process to investigate the time variability of solar photometry and to directly compare simultaneous photometry with radial velocity data. We find that the hyperparameter posteriors are relatively stable over 70 yr of solar photometry and the amplitude tracks the solar cycle. We observe good agreement between the hyperparameter posteriors for the simultaneous photometry and radial velocity data. Our primary conclusion is a recommendation to include an additional prior in Gaussian process fits to constrain the evolutionary timescale to be greater than the recurrence timescale (i.e., the rotation period) to recover more physically plausible and useful results. Our results indicate that such simultaneous monitoring may be a useful tool in enhancing the precision of radial velocity surveys.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2021

Hierarchical Inference with Bayesian Neural Networks: An Application to Strong Gravitational Lensing

In the past few years, approximate Bayesian Neural Networks (BNNs) have demonstrated the ability to produce statistically consistent posteriors on a wide range of inference problems at unprecedented speed and scale. However, any disconnect between training sets and the distribution of real-world objects can introduce bias when BNNs are applied to data. This is a common challenge in astrophysics and cosmology, where the unknown distribution of objects in our universe is often the science goal. In this work, we incorporate BNNs with flexible posterior parameterizations into a hierarchical inference framework that allows for the reconstruction of population hyperparameters and removes the bias introduced by the training distribution. We focus on the challenge of producing posterior PDFs for strong gravitational lens mass model parameters given Hubble Space Telescope–quality single-filter, lens-subtracted, synthetic imaging data. We show that the posterior PDFs are sufficiently accurate (statistically consistent with the truth) across a wide variety of power-law elliptical lens mass distributions. We then apply our approach to test data sets whose lens parameters are drawn from distributions that are drastically different from the training set. We show that our hierarchical inference framework mitigates the bias introduced by an unrepresentative training set's interim prior. Simultaneously, we can precisely reconstruct the population hyperparameters governing our test distributions. Our full pipeline, from training to hierarchical inference on thousands of lenses, can be run in a day. The framework presented here will allow us to efficiently exploit the full constraining power of future ground- and space-based surveys (https://github.com/swagnercarena/ovejero).

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Hierarchical ensemble Kalman methods with sparsity-promoting generalized gamma hyperpriors

This paper introduces a computational framework to incorporate flexible regularization techniques in ensemble Kalman methods, generalizing the iterative alternating scheme to nonlinear inverse problems. The proposed methodology approximates the maximum a posteriori (MAP) estimate of a hierarchical Bayesian model characterized by a conditionally Gaussian prior and generalized gamma hyperpriors. Suitable choices of hyperparameters yield sparsity-promoting regularization. We propose an iterative algorithm for MAP estimation, which alternates between updating the unknown with an ensemble Kalman method and updating the hyperparameters in the regularization to promote sparsity. Here, the effectiveness of our methodology is demonstrated in several computed examples, including compressed sensing and subsurface flow inverse problems.

Ensemble Kalman methods↗
NASA NTRSDate not supplied

Use of Design of Experiments and Rule-Based Inference in Determining Neural Network Architectures for Loss of Control Detection

In this work, we describe methods for selecting the neural network architectures and input spaces to implement belief state inference on generic commercial transport aircraft. First, we highlight a case study on the planning, execution, and analysis of a set of experiments to determine the configurations of a conditional variational autoencoder (CVAE). We present a structured method that can be used in a number of aerospace applications, to optimize the structure and training parameters of the CVAE for belief state inference, using Design of Experiments (DOE) statistical methodologies. The motivation for this specific DOE was to identify the appropriate hyperparameters for measuring the CVAE reconstruction probability and latent space, such that the measurements can be used to infer qualitative state changes for the aircraft. We demonstrate that this process yields information about a trained neural network’s utility for this specific application, along with a quantifiable range of certainty. We execute 84 experiments using loss-of-control flight maneuver data from a NASA T-2 aircraft, demonstrating that this empirical process allows us to construct cheap and simple models with specific attributes amenable to belief state inference in aerospace applications. While theoretically, we could create a single CVAE with an input space the size of all measurable flight variables and environmental dynamics, it becomes intractable to use such a neural network in an in-situ intelligent multi-agent system. Using the recommendations from our case study, we introduce a technical approach for feasibly describing the belief space by (1) identifying significant statistical relationships among flight variables using rule induction, (2) using a set of rules that cover all features to define the input space of multiple CVAEs, and (3) forming a belief space based on the joint probability density of their collective latent spaces. This results in a series of relatively small matrix multiplications that can be performed in real time, as opposed to large matrix computations in a single CVAE. We demonstrate the application of this approach on the T-2 flight loss-of control experiments, using the architecture and hyperparameter recommendations from the case study. We compare the utilities of an individual CVAE trained on all flight variables and multiple CVAEs defined on subsets of flight variables for detecting qualitative changes in flight. We demonstrate that the use of multiple CVAEs with smaller input spaces permits the CVAE to capture more granular relationships in the latent space, permitting better state space characterization and loss-of-control detection.

Design of experiments↗
NASA NTRS2021

Automated Data Accountability for Missions in Mars Rover Data

As the Mars Curiosity Rover transmits data to the JPL Ground Data System (GDS), it frequently observes data loss and corruption, requiring re-transmits from the rover and Ground Data System Analysts (GDSA) to monitor the downlink process. As new missions are launched, the GDSA team redistributes analysts to these new missions, causing shortages in previous missions. The GDSA team can significantly benefit from the automation and optimization of the downlink process of telemetry data. In fact, there is a need for a better understanding of why the data is corrupted, so that the GDSA team can best determine the root cause of the issues in the GDS. This paper presents machine learning and deep learning based approaches to automate and optimize the detection of data loss. We first created a pipeline to automatically accumulate data from the telemetry databases (MAROS, Telemetry Data Storage, and GDS Elastic Search Database) in the downlink process. With our newly created datasets, we perform feature selection to supplement the GDSA understanding of the downlink process and provide supplemental analysis on the importance of different features. We implement various machine learning and deep learning based models, including support vector machines, ensemble methods, and deep neural networks and evaluate their accuracies in identifying whether a downlink process is complete or incomplete. We utilize fast hyperparameter optimization methods that allow our models to quickly be re-trained, allowing them to quickly be tuned and optimized on daily incoming data in real time. This hyperparameter optimization also allows our methods to be quickly integrated into other JPL missions. Our results show that our best-performing machine learning and deep learning based models outperform the existing GDSA detection software by 6 accuracy points and can aid analysts by providing insights into the data accountability problem. Since these various machine learning and deep learning approaches vary significantly in interpretability, we provide a discussion on the tradeoffs between their performance and trustworthiness in helping detect issues in data transmission.

Divsalar, Dariush↗
NASA NTRS2022

Exploring Applications of Machine Learning for Wildfire Monitoring and Detection using Unmanned Aerial Vehicles

Wildfires are increasing in frequency and severity around the world, including the United States. The losses caused by wildfires could be mitigated if high-risk areas, hotspots, and flare-ups could be monitored continuously, such as through the use of Unmanned Aerial Vehicles (UAVs). This paper documents exploratory efforts using machine learning to determine efficient flight paths for UAVs and to detect wildfires using image classification. On path planning, three machine learning techniques—Genetic Algorithm, Simulated Annealing, and Dynamic Programming—were explored. Genetic Algorithm was found to be an effective approach for path planning for wildfire monitoring and surveillance by UAVs. For a scenario of 25 locations in a circular arrangement, the algorithm was able to return the optimal path. The accuracy and execution time was found to be sensitive to the algorithm hyperparameters selected, which was especially evident in scenarios with hundreds or thousands of locations. Simulated Annealing was also found to be an effective approach for UAV path planning, with a major benefit of avoiding getting trapped in local minima and being straightforward to implement. Like Genetic Algorithm, the performance of Simulated Annealing was also found to be sensitive to the algorithm hyperparameters selected. By comparison, Dynamic Programming guarantees optimality for any number of locations, but it was found to be less practical in terms of execution time for scenarios with more than about a couple dozen locations. On wildfire detection, image classification using deep learning with a convolutional neural network was explored. Transfer learning was found to be a useful technique to efficiently train deep learning models. Also, it was determined that GPU processing can increase training speed by an order of magnitude, which enables significantly faster development. For a validation test set of 500 images, there were only two false negatives and zero false positives. These results demonstrate that detecting wildfires in static cameras using machine learning is feasible and establish a baseline for using images captured by UAVs in flight for wildfire detection.

Wildfire management↗
NASA NTRSDate not supplied

Towards an Aviation Large Language Model by Fine-tuning and Evaluating Transformers

In the aviation domain, there are many applications for machine learning and artificial intelligence tools that utilize natural language. For example, there is a desire to know the commonalities in written safety reports such as voluntary post incidents reports or aerial wildfire operations reports to better understand the risks present. Another use-case is the possibility of extracting airspace procedures and constraints currently written in documents such as Letters of Agreement. These applications can benefit from the use of state-of-the-art natural language processing techniques when adapted to the language/phraseology specific to the aviation domain. This paper evaluates the viability of adaptation of NLP tools to the aviation domain by fine-tuning transformer based models using aviation data sets. In 2018, a novel language model based on neural units (also called transformers) was created and became known as “Bidirectional Encoder Representations from Transformers” or BERT. This architecture combined with large amounts of English training data and innovative semi-supervised training tasks set the standard for what would later emerge as Large Language Models. The performance of these models was further improved by hyperparameter tuning and refinement of the semi-supervised training task and resulted in “Robustly Optimized BERT Pre-training Approach through hyperparameter tuning” or RoBERTa models. These pre-trained Large Language Models proved to be useful for a wide variety of natural language processing tasks such as text classification and question answering through a process called fine-tuning. The transformer architecture with pre-trained weights served as the basis with the last few layers replaced with layers fine-tuned to perform a new task e.g., a layer that provides a label for the entire input text. This process of fine-tuning can also be used to adapt the models to new domains; e.g., BioBERT started with the pre-trained BERT model and was completed by additional fine-tuning and training on biomedical documents. Transformer-based architectures can also be used to create rich representations of text called embeddings which can serve as the input to other machine learning models. This allows simpler algorithms such as logistic regression to use context-rich representations of the text while still remaining quick to train and evaluate. In the world of aviation, there is a growing demand for natural language processing and understanding but the domain presents unique challenges. Due to the technical content (and specialized language) of most aviation documents, fine-tuning pre-trained Large Language Models to specific tasks has not met the benchmark on natural language processing tasks set by simpler models trained from scratch on the data. To address this deficiency, this paper evaluates the improvements from fine-tuning a Large Language Model on a large set of aviation documents using the original semi-supervised training tasks before performing specific natural language tasks. In fine-tuning, a domain-specific dataset is used on the original training task but with the pre-trained Large Language Model instead of starting from a random initialization. This approach allows the model to be adapted to the specific domain language without discarding the information gained from training on general English data. This paper utilized two major dataset types to train and assess the RoBERTa fine-tuning performance. The first are 7,057 Letters of Agreement which are Federal Aviation Administration (FAA) documents that formalize airspace operations across the national airspace system. They contain many examples of ‘aviation English’ using domain specific terminology and phrasing which serves as a representative basis to perform the semi-supervised fine-tuning. The second type is the 494 document classification labels to be used for evaluation. This down-stream evaluation aims to show the performance of the fine-tuned model, better understand how much data is needed for an effective fine-tuning, and how fine-tuning can be adapted for different applications in-the domain. After semi-supervised training, evaluation begins by encoding the documents for classification using the fine-tuned RoBERTa model. Then a logistic regression classifier is trained to label the document type and compared against our ground truth labels. This currently leads to a 82.8% accuracy on 10-fold cross validation showing improvement over baseline RoBERTa which achieved 81.0%. We plan to measure the improvements on additional tasks and it is expected that these improvements will lead to more robust models that can tackle the natural language processing challenges present in aviation datasets.

ATM↗
DOE OSTI2022

PI3NN: Out-of-distribution-aware Prediction Intervals from Three Neural Networks

We propose a novel prediction interval (PI) method for uncertainty quantification, which addresses three major issues with the state-of-the-art PI methods. First, existing PI methods require retraining of neural networks (NNs) for every given confidence level and suffer from the crossing issue in calculating multiple PIs. Second, they usually rely on customized loss functions with extra sensitive hyperparameters for which fine tuning is required to achieve a well-calibrated PI. Third, they usually underestimate uncertainties of out-of-distribution (OOD) samples leading to over-confident PIs. Our PI3NN method calculates PIs from linear combinations of three NNs, each of which is independently trained using the standard mean squared error loss. The coefficients of the linear combinations are computed using root-finding algorithms to ensure tight PIs for a given confidence level. We theoretically prove that PI3NN can calculate PIs for a series of confidence levels without retraining NNs and it completely avoids the crossing issue. Additionally, PI3NN does not introduce any unusual hyperparameters resulting in a stable performance. Furthermore, we address OOD identification challenge by introducing an initialization scheme which provides reasonably larger PIs of the OOD samples than those of the in-distribution samples. Benchmark and real-world experiments show that our method outperforms several state-of-the-art approaches with respect to predictive uncertainty quality, robustness, and OOD samples identification.

Liu, Siyan↗