Search NASA⌕ Search

SEARCH · Search NASA

Results for “Pre-trained models”

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 127 records · Page 7
All sourcesNASA NTRSDOE OSTI
DOE OSTI2025

Transfer learning nonlinear plasma dynamic transitions in low dimensional embeddings via deep neural networks

Deep learning algorithms provide a new paradigm to study high-dimensional dynamical behaviors, such as those in fusion plasma systems. Development of novel, data-driven model reduction methods, coupled with detection of abnormal modes with plasma physics, opens a unique opportunity to identify plasma instabilities through automated construction of parsimonious models that can be tuned to balance accuracy and cost. Our fusion transfer learning (FTL) model demonstrates success in rapidly reconstructing nonlinear kink mode structures by learning from a limited amount of nonlinear simulation data. The knowledge transfer process leverages a pre-trained neural encoder–decoder network, initially trained on linear simulations, to effectively capture nonlinear dynamics. The low-dimensional embeddings extract the coherent structures of interest, while preserving the inherent dynamics of the complex system. Experimental results highlight FTL’s capacity to capture transitional behaviors and dynamical features in plasma dynamics—a task often challenging for conventional methods. The model developed in this study is generalizable and can be extended broadly through transfer learning to address various magnetohydrodynamics modes.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2025

Quantum AI Based Enhanced Detection of Dementia

Quantum computing has the potential to significantly improve the early detection of Alzheimer's Disease and Related Dementias (ADRD). Quantum-enhanced machine learning can be used to perform an early screening of Alzheimer's disease using brain imaging data based on dataset of MRI scans from both healthy individuals and those diagnosed with Alzheimer's. This study aims to demonstrate the potential of quantum transfer learning to enhance the performance of the classical deep learning model for dementia detection. Using the MRI sagittal images available in the OASIS-2 (64 demented and 72 non-demented subjects between 60 and 96 years), we show how quantum techniques can transform a suboptimal classical model into a more effective solution for dementia detection, highlighting their potential impact on advancing healthcare technology. We begin with a simple classical deep learning model with a significantly smaller number of parameters, which gives suboptimal performance on the problem. Then, we apply different configurations of quantum transfer learning based on the pre-trained weak classifier (Figure 1). We fix the weak classifier's initial convolutional layers at their fixed pre-trained parameters and replace the last set of dense layers with a dressed quantum circuit (DQN), which we train to enhance performance. We performed 4-fold cross-validation for both the classical and the hybrid quantum models and trained them using Pennylane's `default.qubit' simulator and IonQ's Aria-1 simulator (noisy simulation). We showed that with significantly fewer parameters, the quantum transfer learning-based hybrid models showed significant performance enhancement over the base weak classical deep learning model for dementia detection. To classify between a demented and non-demented subject, the accuracy of quantum-based AI methods improved by 6 to 14% compared to classical methods. The sensitivity of the models improved by 4 to 17%. This shows that there are fewer chances of misclassifying demented patients. Figure 2 compares the performance of the hybrid quantum models and their base classical model, and Table 1 summarizes the results. We illustrated that with assistance from quantum machine learning, it is possible to enhance detection for dementia based on brain images. This shows the potential for practical utility of quantum computing in ADRD research.

Bhowmik, Sounak [University of Tennessee, Knoxvill↗
DOE OSTI2025

DONUT: physics-aware machine learning for real-time X-ray nanodiffraction analysis

Coherent X-ray scattering techniques are critical for investigating the fundamental structural properties of materials at the nanoscale. While advancements have made these experiments more accessible, real-time analysis remains a significant bottleneck, often hindered by artifacts and computational demands. In scanning X-ray nanodiffraction microscopy, which is widely used to spatially resolve structural heterogeneities, this challenge is compounded by the convolution of the divergent beam with the sample’s local structure. To address this, we introduce DONUT (Diffraction with Optics for Nanobeam by Unsupervised Training), a physics-aware neural network designed for the rapid and automated analysis of nanobeam diffraction data. By incorporating a differentiable geometric diffraction model directly into its architecture, DONUT learns to predict crystal lattice strain and orientation in real-time. Crucially, this is achieved without reliance on labeled datasets or pre-training, overcoming a fundamental limitation for supervised machine learning in X-ray science. We demonstrate experimentally that DONUT accurately extracts all features within the data over 200 times more efficiently than conventional fitting methods.

Materials science↗
DOE OSTI2025

DONUT: Physics-aware Machine Learning for Real-time X-ray Nanodiffraction Analysis

SF-25-088 Coherent X-ray scattering techniques are critical for investigating the fundamental structural properties of materials at the nanoscale. While advancements have made these experiments more accessible, real-time analysis remains a significant bottleneck, often hindered by artifacts and computational demands. In scanning X-ray nanodiffraction microscopy, which is widely used to spatially resolve structural heterogeneities, this challenge is compounded by the convolution of the divergent beam with the sample’s local structure. To address this, we introduce DONUT (Diffraction with Optics for Nanobeam by Unsupervised Training), a physics-aware neural network designed for the rapid and automated analysis of nanobeam diffraction data. By incorporating a differentiable geometric diffraction model directly into its architecture, DONUT learns to predict crystal lattice strain and orientation in real-time. Crucially, this is achieved without reliance on labeled datasets or pre-training, overcoming a fundamental limitation for supervised machine learning in X-ray science. We demonstrate experimentally that DONUT accurately extracts all features within the data over 200 times more efficiently than conventional fitting methods.

Zhou, Tao [Argonne National Laboratory (ANL), Argo↗
NASA NTRSDate not supplied

The Application of Remote Sensing and Machine Learning to Improve Early Warning Systems for Harmful Algal Events in the Highland Lake Chain, TX

Beginning in 2019, harmful algal events have caused canine deaths in both Lady Bird Lake and Lake Travis located near Austin, Texas. These two reservoirs are part of the larger Highland Lakes chain, managed by the City of Austin Department of Watershed Protection (COA DWP) and the Lower Colorado River Authority (LCRA), which fulfill municipal, commercial, and agricultural water demands. Given the recent increase in favorable environmental conditions for algal events in central Texas, NASA DEVELOP partnered with LCRA and COA DWP to improve monitoring and early detection of algal events, utilizing satellite remote sensing and machine learning. Spatially and temporally varied chlorophyll a concentrations, cyanobacteria detections, turbidity, and water surface temperature products are used as environmental proxies. Landsat 8 Operational Land Imager (OLI) and the Sentinel-2 MultiSpectral Instrument (MSI) data were used to provide a combined revisit time of up to ~3 days and < 30 m per pixel products. Chlorophyll a concentrations were estimated using a pre-trained Mixture Density Network, and cyanobacteria detection was accomplished using the Broad Wavelength Algae Index, which can differentiate algal blooms from algal proliferations (mats). In situ data were used to validate remotely sensed measurements and quantify uncertainties. Preliminary results show a good fit between the modeled output and in situ observations, suggesting that remote sensing data can be used to retrieve biogeochemical properties and/or inherent optical properties (IOPs) of water columns in these inland human-made lakes. Uncertainties were introduced from the sensitivity to atmospheric correction, inherent mismatch between satellite and sampling data, and a relatively lower signal-to-noise ratio over water. The resulting products enable near real-time monitoring of environmental proxies relevant to algal event presence in the Highland Lakes chain, and will ultimately support water management, decision making, and risk communication.

Shuyu Chang↗
DOE OSTI2025

Benchmarking universal machine learning interatomic potentials for rapid analysis of inelastic neutron scattering data

The accurate calculation of phonons and vibrational spectra remains a significant challenge, requiring highly precise evaluations of interatomic forces. Traditional methods based on the quantum description of the electronic structure, while widely used, are computationally expensive and demand substantial expertise. Emerging universal machine learning interatomic potentials (uMLIPs) offer a transformative alternative by employing pre-trained neural network surrogates to predict interatomic forces directly from atomic coordinates. This approach dramatically reduces computation time and minimizes the need for technical knowledge. In this paper, we produce a phonon database comprising nearly 5000 inorganic crystals to benchmark the performance of several leading uMLIPs. We further assess these models in real-world applications by using them to analyze experimental inelastic neutron scattering data collected on a variety of materials. Through detailed comparisons, we identify the strengths and limitations of these uMLIPs, providing insights into their accuracy and suitability for fast calculations of phonons and related properties, as well as the potential for real-time interpretation of neutron scattering spectra. Our findings highlight how the rapid advancement of AI in science is revolutionizing experimental research and data analysis.

inelastic neutron scattering↗
DOE OSTI2025

You Only Look Once v5 and Multi-Template Matching for Small-Crack Defect Detection on Metal Surfaces

This paper compares the performance of Deep Learning (DL) and multi-template matching (MTM) models for detecting small defects. DL models extract distinguishing features of objects but require a large dataset of images. In contrast, alternative computer vision techniques like MTM need a relatively small dataset. The lack of large datasets for small metal-surface defects has inhibited the adoption of automation in small-defect detection in remanufacturing settings. This motivated this preliminary study to compare template-based approaches, like MTM, with feature-based approaches, such as DL models, for small-defect detection on an initial laboratory and remanufacturing industry dataset. This study used You Only Look Once v5 (YOLOv5) as the DL model and compared its performance against the MTM model for small-crack detection. The findings of our preliminary investigation are as follows: (i) YOLOv5 demonstrated higher performance than MTM in detecting small cracks; (ii) an extra-large variant of YOLOv5 outperformed a small-size variant; (iii) the size and object variety of the data are crucial in achieving robust pre-trained weights for use in transfer learning; and (iv) enhanced image resolution contributes to precise object detection.

32 ENERGY CONSERVATION, CONSUMPTION, AND UTILIZATI↗
DOE OSTI2025

Cross-functional transferability in foundation machine learning interatomic potentials

The rapid development of foundation potentials (FPs) in machine learning interatomic potentials demonstrates the possibility for generalizable learning of the universal potential energy surface. The accuracy of FPs can be further improved by bridging the model from lower-fidelity datasets to high-fidelity ones. In this work, we analyze the challenge of this transfer learning (TL) problem within the CHGNet framework. We show that significant energy scale shifts and poor correlations between GGA and r 2 SCAN hinder cross-functional transferability. By benchmarking different TL approaches on the MP-r 2 SCAN dataset, we demonstrate the importance of elemental energy referencing in the TL of FPs. By comparing the scaling law with and without the pre-training on a low-fidelity dataset, we show that significant data efficiency can still be achieved through TL, even with a target dataset of sub-million structures. We highlight the importance of proper TL and multi-fidelity learning in creating next-generation FPs on high-fidelity data.

Huang, Xu [University of California, Berkeley, CA ↗
NASA NTRSDate not supplied

LSKnowledge: Nexus for Transformative Scientific Discoveries and Enhanced Information Retrieval in NASA Life Sciences Portal

We stand at the brink of an extraordinary transformation in the field of AI, driven by the convergence of generative AI and semantic technologies (e.g., knowledge graphs). This fusion holds immense potential and could redefine the future of scientific exploration, particularly in the realm of life sciences research. In this context, we shed light on the pivotal roles that Large Language Models (LLMs) and semantic technologies will play in advancing research, unearthing and comprehending life sciences information through innovative approaches, and empowering researchers to extract insights from NASA's extensive Life Sciences Data Archive. Within the NASA Life Sciences Portal (NLSP), the integration of LLMs and semantic technologies unlocks several advanced capabilities. First and foremost, it equips scientists with sophisticated tools to manage the ever-expanding wealth of scientific literature and data. Furthermore, it facilitates the creation of knowledge graphs that visually represent intricate relationships among biological entities, enabling comprehensive systems-level analysis. Additionally, the fusion of generative AI (including LLMs) and semantic technology can significantly benefit NASA's life sciences research by enhancing information retrieval and hypothesis generation. These tools enhance natural language understanding, facilitating knowledge discovery within NLSP. The overarching vision is to establish a cohesive knowledge ecosystem within NLSP, harnessing the power of LLMs and semantic technologies to synthesize and cross-reference data from diverse missions, disciplines, and research domains. This holistic approach ultimately deepens our understanding of how space environments impact life sciences data. To advance this initiative, we have launched LSKnowledge, aimed at enhancing the information retrieval capabilities of NLSP. In the short term, our primary goal is to develop a robust semantic search system. This system will empower HRP (Human Research Program) researchers to navigate NLSP data repositories more efficiently and precisely, catalyzing the process of hypothesis formation and scientific breakthroughs. To achieve this, we have employed pre-trained LLMs as part of a semantic search tool that can rank and highlight the most relevant records for user queries. To assess the tool's performance, we have curated a set of approximately 200 queries from subject matter experts (SMEs) and manually ranked the top records retrieved by both the current search system and the new semantic search, using SME judgments as the gold standard for relevancy. Herein, we present the results of our comparative analysis and illustrate how these findings have informed the fine-tuning of the system for enhanced performance. In the long term, our objectives include 1) retrieving publicly available information and integrating it with NLSP data to provide more precise answers to user queries, and 2) incorporating non-textual information from the NLSP database into our approach. In conclusion, the fusion of LLMs and semantic technologies within NLSP represents a pioneering stride towards reshaping the landscape of scientific discovery. This synergy not only equips researchers with powerful tools to navigate the burgeoning sea of information but also facilitates a deeper understanding of complex biological relationships, all while accelerating hypothesis generation and knowledge discovery. Through our initiative, LSKnowledge, we are committed to continually refining and expanding these capabilities, with the aim of not only enhancing information retrieval but also integrating diverse data sources to provide more precise insights. In the grand vision, NLSP strives to become the cornerstone of a comprehensive knowledge ecosystem, unraveling the enigmatic intricacies of life sciences phenomena in the context of space environments.

Life Sciences↗
DOE OSTI2026

Self-supervised and multi-fidelity learning for extended predictive soil spectroscopy

Infrared spectroscopy is a cost-effective, non-destructive, and environmentally benign technology that is increasingly recognized as an important solution for meeting the global demand for soil data. While both near-infrared (NIR) and mid-infrared (MIR) diffuse reflectance spectroscopy enable rapid estimation of soil properties, they present a significant trade-off: NIR offers superior scalability and lower operational costs, whereas MIR provides higher analytical fidelity by capturing fundamental molecular vibrations. In this study, we propose a self-supervised, multi-fidelity learning framework designed to bridge this gap. Our approach leverages large-scale MIR spectral libraries to learn a compact, transferable latent representation, into which NIR spectra are subsequently aligned for downstream prediction. The workflow consists of pretraining a latent model on a large MIR library, adapting the representation using a smaller paired NIR–MIR dataset, and evaluating generalization on an independent external test set. Across a range of chemical and physical soil properties, we found that MIR-derived embeddings improved prediction accuracy relative to baseline models that used raw MIR inputs. Predictions derived from the spectrum conversion (NIR to MIR) task did not match the performance of the original MIR spectra but were similar or superior to predictive performance of NIR-only models, suggesting the unified spectral latent space can effectively leverage the larger and more diverse MIR dataset for prediction of soil properties not well represented in current NIR libraries.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Extracting Material Property Measurements from Scientific Literature with Limited Annotations

Extracting material property data from scientific text is pivotal for advancing data-driven research in chemistry and materials science; however, the extensive annotation effort required to produce training data for named entity recognition (NER) models for this task often makes it a barrier to extracting specialized data sets. Here, in this work, we present a comparative study of the conventional, supervised NER methodology to alternative few-shot learning architectures and large language model (LLM)-based approaches that mitigate the need to label large training data sets. We find that the best-performing LLM (GPT-4o) not only excels in directly extracting relevant material properties based on limited examples but also enhances supervised learning through data augmentation. We supplement our findings with error and data quality assessments to provide a nuanced understanding of factors that impact property measurement extraction.

36 MATERIALS SCIENCE↗
DOE OSTI2025

Efficient online quantum circuit learning with no upfront training

Optimization is a promising candidate for studying the utility of variational quantum algorithms (VQAs). However, evaluating cost functions using quantum hardware introduces runtime overheads that limit exploration. Surrogate-based methods can reduce calls to a quantum computer, yet existing approaches require hyperparameter pre-training and have been tested only on small problems. Here, we show that surrogate-based methods can enable successful optimization at scale, without pre-training, by using radial basis function interpolation (RBF) to construct an adaptive, hyperparameter-free surrogate. Using the surrogate as an acquisition function drives hardware queries to the vicinity of the true optima. For 16-qubit random 3-regular Max-Cut instances with the Quantum Approximate Optimization Algorithm (QAOA), our method outperforms state-of-the-art approaches, without considering their upfront training costs. Furthermore, we successfully optimize QAOA circuits for 127-qubit random Ising models on an IBM processor using 10 4 −10 5 measurements. Strong empirical performance demonstrates the promise of automated surrogate-based learning for large-scale VQA applications.

97 MATHEMATICS AND COMPUTING↗
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 OSTI2026

Discovering Strong Gravitational Lenses in the Dark Energy Survey with Interactive Machine Learning and Crowd-sourced Inspection with Space Warps

We conduct a search for strong gravitational lenses in the Dark Energy Survey (DES) Year 6 imaging data. We implement a pre-trained Vision Transformer (ViT) for our machine learning (ML) architecture and adopt interactive machine learning to construct a training sample with multiple classes to address common types of false positives. Our ML model reduces ∼236 million DES cutout images to 22,564 targets of interest, including ∼85% of previously reported galaxy–galaxy lens candidates discovered in DES. These targets were visually inspected by citizen scientists, who ruled out ∼90% as false positives. Of the remaining 2618 candidates, 149 were expert-classified as “definite” lenses and 516 as “probable” lenses, for a total of 665 systems, with 147 of these candidates being newly identified. Additionally, we trained a second ViT to find double-source plane lens systems, finding at least one double-source system. Our main ViT excels at identifying galaxy–galaxy lenses, consistently assigning high scores to candidates with high expert assessments. The top 800 ViT-scored images include ∼100 of our “definite” lens candidates. This selection is an order of magnitude higher in purity than previous convolutional neural-network-based lens searches and demonstrates the feasibility of applying our methodology for discovering large samples of lenses in future surveys.

79 ASTRONOMY AND ASTROPHYSICS↗
NASA NTRSDate not supplied

Training Airline Pilots for Improved Flight Path Monitoring: The Sensemaking Model Framework

The importance and benefit of improved monitoring is increasingly recognized. Improved training may be a valuable intervention. Our study (conducted 2019) assessed and trained airline First Officers on flight path monitoring skills. The exploratory study assessed monitoring pre-training in a simulator session that included monitoring challenges (8 or 7 events). A 1-hour interactive training followed, based on the Sensemaking Model of Monitoring; it presented concepts and examples using a slide deck, discussion, and simple activities. Post-training assessment used scenarios with analogous monitoring challenges (7 or 8 events) but a different setting. Performance showed significant and relatively consistent improvement. Training monitoring as sensemaking merits further investigation.

training↗
NASA NTRSDate not supplied

How Autonomous Intelligent Systems Can Facilitate Earth-independent Medical Care: Going Beyond Telepresence

During the last decade, teleoperated robotic systems have extended humans’ sensorimotor competence to digitally fly beyond the physical barrier of distance and scale and thus transmit sensorimotor skills of the human through direct communication. Telepresence capabilities have enabled tele-physical remote access at small scales thanks to telerobotic mediums. Although the concept was initially motivated by space applications, such technologies quickly have expanded into the medical domain and resulted in teleoperated medical robots, including telerobotic surgical systems (such as the da Vinci surgical system). Effective telepresence fundamentally depends on an agile, reliable, and secure communication medium that can transmit real-time information between the operator and a remote device. However, direct telepresence may not be achievable for long-duration exploration spaceflight missions. Thus, autonomous systems and local intelligence represent potential solutions to the aforementioned issues. One example solution employs demonstration systems which enable learning from the pre-captured inputs of a skilled human operator. These will be computationally modeled and later probabilistically replicated toward the completion of remote physical tasks when direct telepresence is not viable - such as under communication blackout conditions. In other words, trained autonomous systems (e.g., robots) can perform remote operations that mimic the physical performance of experts during remote operations/training. Beyond learning the physics of the task, autonomous agents can also be used to conduct algorithmic decision-making that mimics the higher-level cognition of the expert. Thus, using an autonomous system, pre-trained cognitive and manipulation-based skills can be leveraged (acquired during pre-mission events) to produce digital twins of an intelligent operator. Such systems can be used for the real-time conduction of intricate tasks in complex and unstructured environments. Such systems will operationalize “cognitive digital twins” and can expand the reach of human cognition and manipulation through the power of data-driven learning from demonstration algorithms. This system category will be discussed as a fully autonomous operation in this talk. In addition to the above, we will also propose and discuss the possibility of partial-automation using remote intelligence and remote sensing. In contrast to full automation, partial automation can close the loop through a local operator equipped with augmented sensory awareness through wearable systems. Such technologies will allow the local operator to conduct delicate tasks while being guided using sensory augmentation and being monitored to gauge her/his level of cognitive focus and performance. The difference with the previous category is that a remote human will conduct the task. Further, rather than making a digital twin of human cognition, we will augment the control inputs of the local human to match those of the skilled expert operator who is not accessible in real-time. Going beyond classic telepresence and thus approaching intelligent telepresence, our vision is that autonomous agents will eventually enable the safe, consistent and efficient delivery of complex, remote and smart medical care during space exploration across operators in an Earth-independent fashion. We will discuss our collective vision from NASA and MERIIT@NYU lab in this talk.

Telepresence↗
NASA NTRS2024

Enhancing Metal Additive Manufacturing Training with the Advanced Vision Language Model: A Pathway to Immersive Augmented Reality Training for Non-Experts

This paper introduces an innovative training system for the Renishaw AM400 metal printer, leveraging the synergy of the advanced Vision Language Model (VLM) with Augmented Reality (AR) within the Digital Twins (DT) framework. Aimed at overcoming the limitations of conventional training methods in metal additive manufacturing (AM), our system integrates AR to provide an immersive learning environment, enhancing the real-world experience with interactive digital overlays. The core of the system lies in its use of VLM, which, pre-trained on diverse datasets, excels in processing multi-modal data, thereby offering nuanced and contextually relevant guidance for trainees. Key experiments demonstrate the system’s effectiveness, particularly highlighting the usage of VLM as an Artificial Intelligence (AI) agent to integrate external tools like YOLO-v7 for valve state classification and CRAFT for control panel text recognition. This approach significantly improves recognition accuracy, operational understanding, and human–machine interaction, especially for non-expert users, making complex metal AM operations more accessible. The research not only showcases the potential of AR and VLM in industrial training but also sets a new standard for smart manufacturing practices, indicating broader applications in various industrial domains.

Metal additive manufacturing↗
DOE OSTI2024

Evaluating the Use of Foundational Chemical Language Models in Multimodal Graph Fusion

Rapid and accurate prediction of the physicochemical properties of molecules given their structures remains a key challenge in cheminformatics. Machine learning approaches offer high-throughput options, but the optimality of inductive biases and data representations are up for debate. For example, BERT-based masked language models (MLMs) can be trained in a self-supervised way on hundreds of millions to billions of readily available SMILES strings. Another option is graph neural networks (GNNs), which can operate directly on molecular structures. Yet, generating accurate molecular geometry is computationally expensive, leading to a relative scarcity in data compared to SMILES strings. It is attractive to combine these two paradigms by pre-training an LM on a large corpus of SMILES strings and embedding these representation into a geometric graph neural network. Despite the promise of such an approach, and contrary to previous studies, we find mixed results with the combination of the LMs and GNNs on several molecule datasets. In particular, we found evidence for improvement on the FreeSolv and QM7 benchmarks, but degraded performance on the ESOL, LIPO and QM9 datasets compared to a GNN baseline.

Francel, Collin [University of Alabama]↗