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

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.

💡 Research Summary

The paper tackles the chronic shortage of high‑quality training data for machine‑learning‑driven materials discovery, especially for magnetic properties that are difficult to compute reliably with high‑throughput density‑functional theory. While large computational databases (e.g., Materials Project) provide millions of entries for formation energy or band gap, they lack accurate magnetic descriptors. Conversely, experimental literature contains abundant magnetic measurements, but most curated experimental databases (such as the Northeast Magnetic Materials Database, NEMAD) only store composition and coarse structural tags (e.g., space group), omitting full atomic coordinates. This omission prevents the use of advanced graph‑based neural networks that require precise crystal graphs.

To bridge this gap, the authors develop a two‑stage workflow. First, they align NEMAD entries with crystallographic information files (CIFs) from the Inorganic Crystal Structure Database (ICSD). Matching proceeds in two steps: (i) reduced chemical formula matching using pymatgen, and (ii) space‑group number comparison. When multiple CIFs satisfy the same composition and space group, one is randomly selected. The authors introduce a quantitative “noise” metric ε based on the variance of Niggli‑reduced metric tensors among CIFs matched to the same NEMAD entry; lower ε indicates higher structural consistency. Two aligned datasets are produced: (a) composition‑only alignment (≈44 k Néel and 28 k Curie temperature records) and (b) composition + space‑group alignment (≈5 k Néel and 3.8 k Curie records).

Second, the aligned databases are used to train Crystal Graph Convolutional Neural Networks (CGCNNs). In CGCNN, each atom becomes a node with a 64‑dimensional feature vector, edges encode inter‑atomic distances (and optionally Voronoi‑based geometric descriptors). The model comprises three convolutional layers (128 hidden units each), a graph‑pooling readout, and fully‑connected output layers for regression (temperature) or classification (magnetic ordering). Training follows an 80/10/10 split, Adam optimizer (lr = 0.001) with cosine annealing, early stopping after 20 epochs without validation improvement, and up to 300 epochs.

Performance is evaluated across three alignment conditions (no alignment, composition‑only, composition + space‑group) and two learning strategies (training from scratch, fine‑tuning a CGCNN pre‑trained on formation‑energy data). Metrics include Mean Absolute Error (MAE), R² for regression, and Correct Classification Rate (CCR) for magnetic ordering. Results show a clear trend: stricter alignment reduces ε and improves all metrics. Compared with the original NEMAD models (random forest, XGBoost), CGCNNs trained on the composition‑only aligned set reduce MAE by ~20 % and raise CCR from 0.90 to 0.93. The composition + space‑group set further lowers MAE (Curie ≈ 22.6 K, Néel ≈ 22.0 K) and raises CCR to 0.95. Transfer learning yields additional gains, especially for the smaller composition + space‑group set, where fine‑tuned models achieve the lowest MAE and highest R², demonstrating the value of leveraging a pre‑learned crystal‑representation.

Cross‑validation on the independent MagNData benchmark confirms that models trained on the aligned databases predict experimental temperatures with high fidelity, underscoring that full atomic‑coordinate information is essential for accurate magnetic property prediction.

In summary, the authors contribute (1) an automated pipeline that aligns experimental property records with high‑quality CIFs, (2) a noise metric to assess alignment reliability, and (3) state‑of‑the‑art GNN models that exploit the enriched structural data. The workflow dramatically improves magnetic property predictions and offers a scalable template for other property domains (e.g., catalytic activity, battery electrode behavior), effectively merging literature‑derived experimental data with crystallographic databases to accelerate materials discovery.