PaperlayerView exact revision on arXiv

Source-linked AI summary

Recent Advances and Applications of Deep Learning Methods in Materials Science

Kamal Choudhary, Brian DeCost, Chi Chen, Anubhav Jain, Francesca Tavazza, Ryan Cohn, Cheol WooPark, Alok Choudhary, Ankit Agrawal, Simon J. L. Billinge, Elizabeth Holm, Shyue Ping Ong, Chris Wolverton

arXiv:2110.14820v1cond-mat.mtrl-sciphysics.comp-ph

TL;DR

Materials science links information across atomic, microscale, mesoscale, image, and spectral representations, making cross-component relationships challenging. This review surveys deep-learning principles, recent applications, uncertainty quantification, and evolving resources, concluding that dataset quality, structure sensitivity, and model uncertainty remain important boundaries.

  • Problem

    Linking material information across atomic, microscale, mesoscale, image, and spectral representations is challenging because structures and phenomena span significantly different length and time scales.

  • Method

    The article reviews deep-learning principles and synthesizes major application trends, methods, uncertainty-quantification approaches, limitations, and publicly available resources.

  • Results

    The review identifies deep-learning applications across materials data modalities and provides a GitHub repository intended to evolve as new tools and datasets become publicly available.

  • Takeaways & Limitations

    Deep learning supports materials analysis and autonomous experimentation, including algorithmic adjustment of synthesis conditions based on prior experiments without a human in the loop.

  • Takeaways & Limitations

    Dataset reliability and quality assessment remain difficult because ground truth may be lacking, global comparison metrics may be insufficient, and similar setups may not be reproducible.

Abstract

from arXiv · show

Deep learning (DL) is one of the fastest growing topics in materials data science, with rapidly emerging applications spanning atomistic, image-based, spectral, and textual data modalities. DL allows analysis of unstructured data and automated identification of features. Recent development of large materials databases has fueled the application of DL methods in atomistic prediction in particular. In contrast, advances in image and spectral data have largely leveraged synthetic data enabled by high quality forward models as well as by generative unsupervised DL methods. In this article, we present a high-level overview of deep-learning methods followed by a detailed discussion of recent developments of deep learning in atomistic simulation, materials imaging, spectral analysis, and natural language processing. For each modality we discuss applications involving both theoretical and experimental data, typical modeling approaches with their strengths and limitations, and relevant publicly available software and datasets. We conclude the review with a discussion of recent cross-cutting work related to uncertainty quantification in this field and a brief perspective on limitations, challenges, and potential growth areas for DL methods in materials science. The application of DL methods in materials science presents an exciting avenue for future materials discovery and design.

1 Introduction

Materials science spans widely varying structural and phenomenological scales, making relationships between processing, structure, properties, and performance difficult to establish. Deep learning is increasingly used to analyze growing materials datasets and extend computational approaches, but its black-box nature limits physical interpretability.

  • Materials information ranges from atomic coordinates and microscale phase distributions to mesoscale connectivity, images, and spectra.These structural and phenomenological scales vary substantially, complicating links across processing, structure, properties, and performance.
  • The expansion of public experimental and computational datasets creates demand for automated analysis supported by machine learning.Automation, computational resources, the Materials Genome Initiative, and FAIR principles have contributed to rapid dataset growth.
  • Deep learning can extract higher-level features from raw data and has begun complementing physics-based methods in materials design.Transfer learning, multi-fidelity modeling, and active learning can make DL feasible even with small datasets.
  • Materials design is costly and time-consuming experimentally, while computational methods remain constrained by length and time scales.The number of possible material combinations is too large for exhaustive experiments, and methods such as density functional theory and molecular dynamics retain scale limitations.
  • Deep learning’s black-box nature can hinder physical insight, and interpreting models with thousands to millions of parameters remains difficult.Improving interpretability and explainability is identified as an active research area.
  • This review surveys deep learning principles and applications across materials science while tracking evolving tools and datasets through a public repository.The discussion covers recent trends and provides a GitHub repository intended to be updated as new resources become available.

2 Basics of deep learning

Deep learning extends machine learning through multilayer neural networks that learn representations from structured and unstructured materials data. Training combines parameterized transformations, nonlinear activations, loss optimization, normalization, batching, and regularization.

  • Deep learning is a machine-learning subset using multilayer neural networks, while machine learning is itself a subset of artificial intelligence.AI broadly concerns machines mimicking human intelligence; ML learns from data without explicit task-specific programming.
  • Deep learning is especially useful for unstructured pixels, sounds, text, and graph data because it learns representations that traditional models otherwise require users to engineer.Manual feature extraction can be time-consuming, brittle, and difficult to scale.
  • A perceptron computes a weighted sum of inputs and biases, forming a basic building block for artificial neural networks.Common neural-network training packages include PyTorch, TensorFlow, and MXNet.
  • Neural networks transform inputs through weight matrices and nonlinear activations, then optimize a loss comparing predictions with known targets.Gradient-based methods such as SGD and Adam update parameters using task-dependent losses including BCE, NLLL, and MSE.
  • Training typically uses normalized inputs, mini-batches, and separate training, validation, and test sets.Dropout and early stopping are used to control overfitting and improve generalization.

2.3 Convolution neural networks

Convolutional neural networks learn local, translation-invariant features from regular-grid data through shared filters, nonlinearities, spatial coarsening, and prediction layers. Graph neural networks extend deep learning to non-Euclidean structures such as atomic and molecular systems.

  • Convolutional neural networks: CNNs use learnable shared filters to extract local feature maps from regular-grid inputs with a translation-invariant inductive bias.Filters slide across inputs using convolution rather than general matrix multiplication.
  • Convolutional neural networks: CNNs combine convolution filters, nonlinear activations, spatial coarsening, and a prediction module often based on fully connected layers.Pooling downsamples feature maps while preserving spatial invariance and reducing dimensionality.
  • Graph neural networks: Graph-based methods are important when data structures such as atomic or molecular systems are non-Euclidean rather than regular grids.Graphs represent nodes, edges, and associated features or relations.
  • Graph neural networks: Graph neural networks predict materials properties by repeatedly aggregating neighboring information and updating node or edge representations.Neighbor aggregation is permutation invariant, and common variants include GCN, GAT, Relational-GCN, GRN, GIN, and LGNN.

2.5 Sequence-to-sequence models

Sequence-to-sequence models preserve contextual information while mapping variable-length input sequences to variable-length outputs. Transformers address key recurrent-network limitations through attention-based parallel access to prior words.

  • Sequence-to-sequence models: Bag-of-words converts a document into a fixed-length vector by counting words across the vocabulary.This representation does not itself encode sequential context.
  • Sequence-to-sequence models: Sequence-to-sequence models use contextual word order and can produce outputs of arbitrary length.In named entity recognition, a chemical abstract is mapped to a category for every word.
  • Recurrent neural networks: Recurrent neural networks use feedback connections and shared weights across unfolded hidden layers to process sequences.RNNs include gated recurrent unit and related architectures.
  • Recurrent neural networks: RNNs are difficult to parallelize for large datasets and can lose long-range context because of vanishing gradients.Long short-term memory networks have nevertheless been applied to materials-domain named entity recognition.
  • Transformers: Transformer models use attention to access previous words simultaneously, facilitating parallelization and better preserving long-term context.BERT is presented as an example of a transformer-based sequence-to-sequence model.
  • Generative models: Generative models learn underlying data distributions for tasks such as augmentation and latent-feature learning, unlike supervised discriminative models.Examples include variational autoencoders and generative adversarial networks.

3 Applications of DL methods

Successful DL applications in materials science depend on suitable data, representations, and evaluation criteria. The review surveys key application areas built around these considerations.

  • Large, balanced, and diverse datasets are often needed for successful materials-science DL applications.The passage notes that datasets often contain on the order of 10000 data points or more.
  • Choosing an appropriate DL approach and vector or graph representation is a central modeling consideration.
  • Performance metrics should be selected to match the scientific goals of the application.

3.1 Atomistic and chemical representations

Atomistic DL represents materials through structures, compositions, or related descriptors to accelerate force-field development, property prediction, screening, and inverse design. These models can reduce computational cost, but their accuracy and applicability depend on structural information, training data, and representation choice.

  • Atomistic DL addresses force-field development, direct property prediction, materials screening, and generative inverse design.
  • DL force fields can make predictions hundreds to thousands of times faster, but training still requires computationally expensive datasets.
  • Graph neural network force fields extract translationally invariant and rotationally covariant structural features for direct atomic-force prediction.
  • ALIGNN reports mean absolute errors of 0.022 eV/atom for solid formation energy and 0.002 eV for molecular internal energy at 0 K.
  • DL property predictors support rapid prescreening before traditional DFT calculations, which are limited to relatively few compounds and require substantial computation.
  • Full atomistic representations require crystal structures and atom positions that are often obtained only after DFT relaxation or experiment.Alternative representations avoid this requirement, but their errors tend to exceed those of graph-based models.

3.2 Chemical formula and segment representations

Chemical formula, SMILES, and fragment representations provide lower-cost alternatives to full atomistic descriptions for molecular and materials DL. They enable screening and property prediction, while sacrificing some structural detail or validity guarantees.

  • 3.2.1 SMILES and fragment representation: SMILES encodes molecular elements and bonding as ASCII strings, including structural differences such as chirality.
  • 3.2.1 SMILES and fragment representation: SMILES strings can be converted into fixed-length sparse feature matrices and augmented with elemental and bonding attributes for DL.
  • 3.2.1 SMILES and fragment representation: VAE, GAN, and RNN models generate molecules with target properties, although generated SMILES may not correspond to physically valid structures.
  • 3.2.1 SMILES and fragment representation: Fragment-based models such as DeepFrag predict suitable ligand substituents after removing a fragment from a ligand–receptor complex.
  • 3.2.2 Chemical formula representation: Chemical-formula representations range from raw elemental fractions to descriptors incorporating attributes such as electronegativity and valency.
  • 3.2.2 Chemical formula representation: ElemNet significantly outperformed Random Forest using only raw elemental fractions, while learning chemistry-related patterns without explicit periodic-table information.
  • 3.2.2 Chemical formula representation: Composition-based representations support diverse property prediction and transfer learning, including applications to experimental datasets where atomic structures are unavailable.
  • 3.2.2 Chemical formula representation: Composition-based models cannot distinguish polymorphs and accurate formation-energy predictions do not necessarily yield accurate stability predictions.

3.3 Spectral models

Deep learning is increasingly applied to spectral data for structural identification, phase classification, local-environment analysis, and prediction of material properties. Applications use both computationally generated spectra and experimental datasets, with synthetic data supporting several high-accuracy demonstrations.

  • Spectroscopy provides composition, structural, and dynamic information, while growing high-throughput datasets create opportunities for DL-based analysis.
  • Large spectral resources include computational databases, curated experimental collections, and high-throughput datasets such as 50000 optical absorption spectra in HTEM.
  • CNNs trained on 150000 simulated ICSD XRD patterns reached 81.14% space-group, 83.83% extinction-group, and 94.99% crystal-system classification accuracy.
  • DL models classify phases and can estimate compound fractions in multiphase XRD mixtures using large synthetic training sets.
  • XAS models predict coordination numbers, classify topological materials, and infer longer-range radial distribution functions from XANES or transformed EXAFS data.
  • Deep learning models for vibrational spectra have been reported to outperform chemometric analyses in several studies.
  • DOSnet uses DFT electronic density-of-states spectra to predict adsorption energies for adsorbates on bimetallic alloy surfaces.

3.4 Image based models

Materials imaging DL applications span classification, regression, segmentation, superresolution, recognition, tracking, and representation learning across microscopy modalities. These methods reduce manual feature engineering and support automated, high-throughput analysis of atomic- to mesoscale structures.

  • Computer-vision tasks in materials microscopy include classification, regression, hyperparameter tuning, segmentation, superresolution, recognition, tracking, and microstructure representation learning.
  • These tasks generalize across optical, electron, scanning-probe, transmission-electron, and scanning-transmission-electron microscopy modalities.
  • Microscopy images capture atomic- to mesoscale structures, defects, and dynamics linked to material functionality and performance.
  • Conventional image-processing pipelines have advanced quantitative microscopy but are often brittle and require extensive manual feature identification.
  • DL is presented as a promising approach for high-performance, high-throughput automated analysis of materials image datasets.

3.4.1 Databases and software libraries

Materials image-learning datasets can be experimental, simulated, or literature-derived, while software packages support increasingly diverse image-learning tasks. The review catalogs representative databases and packages for this growing application area.

  • Materials image datasets may originate from experiments, simulations, or published literature.Libraries can also generate STM and STEM images computationally.
  • The review presents common materials image datasets and useful software packages for image-learning applications.

3.4.2 Applications

Applications demonstrate DL for materials-image classification, segmentation, defect analysis, superresolution, instrument control, manufacturing monitoring, and autonomous experimentation. Reported studies span simulated and experimental data, with gains in accuracy, throughput, and automation alongside clear task-specific limitations.

  • Image classification and regression: Transfer learning classified diverse SEM material systems with 90 % test accuracy, while mixed-class images caused misclassifications.
  • Image classification and regression: A CNN classified five 2D Bravais lattices from simulated STM images with an average F1 score of around 0.9 for each lattice type.
  • Image classification and regression: DL improved diffraction and EBSD analysis, including high symmetry classification and 16 % lower disorientation error than dictionary-based indexing on experimental patterns.
  • Image classification and regression: A residual DL model classified prior deformation history with 92.48 % accuracy versus 68.24 % for a correlation-function method and also predicted stress-strain curves.
  • Pixel-wise learning: Image segmentation models achieved 94 % accuracy for steel phases and 86.5 % precision and recall for phase segmentation and 91.1 % for spheroidite particles.
  • Object/entity recognition, localization, and tracking: Atomistic models detected defects and atomic positions, including experimental positions after simulation-only training, while a WSe2 U-Net reached up to 98 % accuracy.
  • Object/entity recognition, localization, and tracking: DefectSegNet produced quantitatively more accurate and consistent measurements than manual measurements by multiple human experts.
  • Superresolution imaging and auto-tuning experimental parameters: GAN-based reconstruction recovered high-resolution STEM images from sampling areas reduced by up to 100, while DL superresolution reduced electron-beam interaction from 13.9 % to 3.7 %.

3.5 Natural language processing

Materials NLP converts largely unstructured literature into structured information, supporting information extraction, search, synthesis prediction, and materials discovery. Progress is enabled by specialized preprocessing, embeddings, and deep models, but access and parsing remain substantial constraints.

  • Materials knowledge is often stored as unstructured text, tables, or images, motivating NLP methods that convert publications into structured data or learn directly from text.
  • NLP datasets include journal articles, preprints, patents, and online material, but copyright restrictions, publisher agreements, PDF formats, and parsing difficulties limit access and use.
  • NLP workflows retrieve data, preprocess and featurize text, then apply machine learning for tasks such as entity recognition, relationship modeling, and question answering.
  • Materials-specific preprocessing and embeddings better preserve chemical formulas, scientific notation, chemical entities, and domain meaning than generic text processing.
  • Deep learning models, including LSTMs and transformers, extract materials synthesis information used to predict synthesis maps for titania nanotubes, binary and ternary oxides, and perovskites.
  • NLP-derived databases support materials discovery, including identifying HoBe2's large magnetocaloric properties and screening dye candidates for solar-cell co-sensitization.
  • Word embeddings trained on materials literature can directly predict applications through dot products between composition and application words, without intermediary models.

4 Uncertainty quantification

Uncertainty quantification is presented as essential for evaluating the robustness, reliability, and trustworthiness of deep-learning predictions in materials science. The review surveys common approaches while emphasizing that scalable, broadly applicable uncertainty estimation remains unresolved.

  • Motivation: Uncertainty quantification is essential for evaluating the robustness of deep-learning models and can address concerns about their reliability and interpretability.Determining uncertainty remains challenging and far from completely solved.
  • Limitations: Most uncertainty-quantification implementations do not work with arbitrary off-the-shelf models without retraining or redesigning.Bayesian neural networks avoid this restriction but require substantial training modifications and computational expense.
  • Motivation: Limited and insufficiently diverse materials datasets make high predictive accuracy unreliable, particularly for extrapolation and ambiguous samples.Uncertainty can help estimate required training-set size and identify additional data needs.
  • Methods: Dropout, deep ensembles, quantile regression, and Gaussian processes are the main approaches reviewed for estimating predictive uncertainty.Monte Carlo dropout applies randomized node selection during prediction, while ensembles use multiple models to generate output distributions.
  • Methods: Quantile regression estimates prediction intervals from selected quantiles, with amin = 0.1 and amax = 0.8 producing intervals covering 70 % of the population.The approach traditionally requires three model runs for three quantiles, although TensorFlow can obtain multiple quantiles simultaneously.

5 Limitations and challenges

The review identifies dataset quality, representation, realism of training data, interpretability, evaluation practice, and autonomous experimentation as continuing challenges for deep learning in materials science. These constraints limit generalization, physical insight, and reliable deployment across materials problems.

  • Data and evaluation: Dataset reliability and quality assessment remain difficult because ground truth, global comparison metrics, and reproducibility can be lacking.These issues complicate reliance on deep-learning-based predictions.
  • Representations: Chemical-formula representations omit structure, limiting their ability to capture structure-sensitive phenomena such as phase transitions.Materials with identical formulas can have sharply different properties depending on atomic arrangement, as illustrated by diamond and graphite.
  • Representations: Atomistic graph models have mainly been tested on bulk materials, leaving defective systems and multidimensional phase-space exploration less established.Despite improved accuracy, model errors still need reduction toward deep-learning chemical accuracies.
  • Images and spectra: Experimental images and spectra are often noisy and require substantial preprocessing, whereas simulated data may fail to represent realistic scenarios.This creates a trade-off between experimental complexity and the realism limitations of theory-based training data.
  • Interpretability: Deep-learning models remain difficult to interpret, motivating materials-specific efforts such as GNNExplainer to improve community trust.The review identifies interpretability as an important area for future work.
  • Evaluation: Standard train-validation-test splits may overestimate generalization when similar materials appear in training and test sets.The review calls for better practices in splitting, normalization, and augmentation.
  • Autonomous experimentation: A major technological challenge is integrating machine learning and experiments into a closed-loop autonomous materials design and synthesis process.Such systems are described as self-driving laboratories, with early proof-of-principle attempts noted.
Loading 2110.14820v1…