skip to main content


Title: Towards overcoming data scarcity in materials science: unifying models and datasets with a mixture of experts framework
Abstract

While machine learning has emerged in recent years as a useful tool for the rapid prediction of materials properties, generating sufficient data to reliably train models without overfitting is often impractical. Towards overcoming this limitation, we present a general framework for leveraging complementary information across different models and datasets for accurate prediction of data-scarce materials properties. Our approach, based on a machine learning paradigm called mixture of experts, outperforms pairwise transfer learning on 14 of 19 materials property regression tasks, performing comparably on four of the remaining five. The approach is interpretable, model-agnostic, and scalable to combining an arbitrary number of pre-trained models and datasets to any downstream property prediction task. We anticipate the performance of our framework will further improve as better model architectures, new pre-training tasks, and larger materials datasets are developed by the community.

 
more » « less
Award ID(s):
2118201 2106825
NSF-PAR ID:
10380747
Author(s) / Creator(s):
; ;
Publisher / Repository:
Nature Publishing Group
Date Published:
Journal Name:
npj Computational Materials
Volume:
8
Issue:
1
ISSN:
2057-3960
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Abstract

    Generative models are a sub-class of machine learning models that are capable of generating new samples with a target set of properties. In chemical and materials applications, these new samples might be drug targets, novel semiconductors, or catalysts constrained to exhibit an application-specific set of properties. Given their potential to yield high-value targets from otherwise intractable design spaces, generative models are currently under intense study with respect to how predictions can be improved through changes in model architecture and data representation. Here we explore the potential of multi-task transfer learning as a complementary approach to improving the validity and property specificity of molecules generated by such models. We have compared baseline generative models trained on a single property prediction task against models trained on additional ancillary prediction tasks and observe a generic positive impact on the validity and specificity of the multi-task models. In particular, we observe that the validity of generated structures is strongly affected by whether or not the models have chemical property data, as opposed to only syntactic structural data, supplied during learning. We demonstrate this effect in both interpolative and extrapolative scenarios (i.e., where the generative targets are poorly represented in training data) for models trained to generate high energy structures and models trained to generated structures with targeted bandgaps within certain ranges. In both instances, the inclusion of additional chemical property data improves the ability of models to generate valid, unique structures with increased property specificity. This approach requires only minor alterations to existing generative models, in many cases leveraging prediction frameworks already native to these models. Additionally, the transfer learning strategy is complementary to ongoing efforts to improve model architectures and data representation and can foreseeably be stacked on top of these developments.

     
    more » « less
  2. Abstract

    Predicting properties from a material’s composition or structure is of great interest for materials design. Deep learning has recently garnered considerable interest in materials predictive tasks with low model errors when dealing with large materials data. However, deep learning models suffer in the small data regime that is common in materials science. Here we develop the AtomSets framework, which utilizes universal compositional and structural descriptors extracted from pre-trained graph network deep learning models with standard multi-layer perceptrons to achieve consistently high model accuracy for both small compositional data (<400) and large structural data (>130,000). The AtomSets models show lower errors than the graph network models at small data limits and other non-deep-learning models at large data limits. They also transfer better in a simulated materials discovery process where the targeted materials have property values out of the training data limits. The models require minimal domain knowledge inputs and are free from feature engineering. The presented AtomSets model framework can potentially accelerate machine learning-assisted materials design and discovery with less data restriction.

     
    more » « less
  3. Many applications of machine learning require a model to make accurate predictions on test examples that are distributionally different from training ones, while task-specific labels are scarce during training. An effective approach to this challenge is to pre-train a model on related tasks where data is abundant, and then fine-tune it on a downstream task of interest. While pre-training has been effective in many language and vision domains, it remains an open question how to effectively use pre-training on graph datasets. In this paper, we develop a new strategy and self-supervised methods for pre-training Graph Neural Networks (GNNs). The key to the success of our strategy is to pre-train an expressive GNN at the level of individual nodes as well as entire graphs so that the GNN can learn useful local and global representations simultaneously. We systematically study pre-training on multiple graph classification datasets. We find that naïve strategies, which pre-train GNNs at the level of either entire graphs or individual nodes, give limited improvement and can even lead to negative transfer on many downstream tasks. In contrast, our strategy avoids negative transfer and improves generalization significantly across downstream tasks, leading up to 9.4% absolute improvements in ROC-AUC over non-pre-trained models and achieving state-of-the-art performance for molecular property prediction and protein function prediction. 
    more » « less
  4. Abstract Motivation

    Properties of molecules are indicative of their functions and thus are useful in many applications. With the advances of deep-learning methods, computational approaches for predicting molecular properties are gaining increasing momentum. However, there lacks customized and advanced methods and comprehensive tools for this task currently.

    Results

    Here, we develop a suite of comprehensive machine-learning methods and tools spanning different computational models, molecular representations and loss functions for molecular property prediction and drug discovery. Specifically, we represent molecules as both graphs and sequences. Built on these representations, we develop novel deep models for learning from molecular graphs and sequences. In order to learn effectively from highly imbalanced datasets, we develop advanced loss functions that optimize areas under precision–recall curves (PRCs) and receiver operating characteristic (ROC) curves. Altogether, our work not only serves as a comprehensive tool, but also contributes toward developing novel and advanced graph and sequence-learning methodologies. Results on both online and offline antibiotics discovery and molecular property prediction tasks show that our methods achieve consistent improvements over prior methods. In particular, our methods achieve #1 ranking in terms of both ROC-AUC (area under curve) and PRC-AUC on the AI Cures open challenge for drug discovery related to COVID-19.

    Availability and implementation

    Our source code is released as part of the MoleculeX library (https://github.com/divelab/MoleculeX) under AdvProp.

    Supplementary information

    Supplementary data are available at Bioinformatics online.

     
    more » « less
  5. Abstract Modern machine learning (ML) and deep learning (DL) techniques using high-dimensional data representations have helped accelerate the materials discovery process by efficiently detecting hidden patterns in existing datasets and linking input representations to output properties for a better understanding of the scientific phenomenon. While a deep neural network comprised of fully connected layers has been widely used for materials property prediction, simply creating a deeper model with a large number of layers often faces with vanishing gradient problem, causing a degradation in the performance, thereby limiting usage. In this paper, we study and propose architectural principles to address the question of improving the performance of model training and inference under fixed parametric constraints. Here, we present a general deep-learning framework based on branched residual learning (BRNet) with fully connected layers that can work with any numerical vector-based representation as input to build accurate models to predict materials properties. We perform model training for materials properties using numerical vectors representing different composition-based attributes of the respective materials and compare the performance of the proposed models against traditional ML and existing DL architectures. We find that the proposed models are significantly more accurate than the ML/DL models for all data sizes by using different composition-based attributes as input. Further, branched learning requires fewer parameters and results in faster model training due to better convergence during the training phase than existing neural networks, thereby efficiently building accurate models for predicting materials properties. 
    more » « less