A New Workflow for Materials Discovery Bridging the Gap Between Experimental Databases and Graph Neural Networks

TL;DR

The paper tackles data scarcity in ML-guided materials discovery by bridging experimental magnetic-material data with complete crystallographic coordinates through alignment of NEMAD against ICSD CIFs. It introduces two CIF-alignment policies to create structure-complete databases, evaluates them with CGCNNs in scratch and transfer-learning modes, and develops a noise metric to quantify alignment ambiguity. Results show substantial reductions in $MAE$ and improvements in $CCR$ for Néel and Curie temperature predictions, with additional gains from pretraining, though domain mismatches can limit transfer learning on magnetic benchmarks like MagNData. This workflow enables more accurate structure-aware predictions and offers a scalable path toward accelerated discovery of magnetic materials and related properties.

Abstract

Incorporating Machine Learning (ML) into material property prediction has become a crucial step in accelerating materials discovery. A key challenge is the severe lack of training data, as many properties are too complicated to calculate with high-throughput first principles techniques. To address this, recent research has created experimental databases from information extracted from scientific literature. However, most existing experimental databases do not provide full atomic coordinate information, which prevents them from supporting advanced ML architectures such as Graph Neural Networks (GNNs). In this work, we propose to bridge this gap through an alignment process between experimental databases and Crystallographic Information Files (CIF) from the Inorganic Crystal Structure Database (ICSD). Our approach enables the creation of a database that can fully leverage state-of-the-art model architectures for material property prediction. It also opens the door to utilizing transfer learning to improve prediction accuracy. To validate our approach, we align NEMAD with the ICSD and compare models trained on the resulting database to those trained on NEMAD originally. We demonstrate significant improvements in both Mean Absolute Error (MAE) and Correct Classification Rate (CCR) in predicting the ordering temperatures and magnetic ground states of magnetic materials, respectively.

Figures (3)

• Figure 1: Schematic representing our alignment process followed by subsequent training of CGCNNs. As can be seen above, the database aligned only on composition is prone to align CIFs with incorrect space groups to NEMAD entries.
• Figure 2: Plots of the ground truth Néel/Curie temperature vs. model prediction for models trained from scratch on databases aligned with composition alone, and aligned with both composition and space group. For Néel temperature predictions, we also include the results from the models fine-tuned from those trained on formation energy.
• Figure 3: Plots of ground truth Néel temperature vs. model prediction on the MagNData database using models trained from scratch and fine-tuned from those trained on formation energy, along with the MedAE value for each. We show results for both models trained on the database aligned only by composition, and models trained on the database aligned by composition and space group.