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

Brandon Schoener, Yuting Hu+7 more|arXiv (Cornell University)|Jan 31, 2026

Machine Learning in Materials Science0 citations

TL;DR

The paper aligns experimental magnetic-material data with CIF structures from ICSD to train CGCNNs for Curie/Néel temperatures, showing improved MAE and CCR over composition-only alignment and enabling transfer learning.

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.

Motivation & Objective

• Motivate the use of ML for rapid magnetic-material property prediction when high-quality training data are scarce.
• Propose a workflow that aligns experimental databases with CIFs from ICSD to create complete-structure datasets for ML.
• Demonstrate that CGCNNs trained on aligned data outperform composition-only models in predicting magnetic properties.
• Show that transfer learning from related tasks further improves prediction accuracy.

Proposed method

• Retrieve ICSD CIFs and NEMAD composition entries.
• Match CIFs to NEMAD entries by reduced chemical formula and space group to create aligned databases.
• Quantify alignment ambiguity with a noise metric based on metric tensors of Niggli-reduced cells.
• Train CGCNNs from scratch on the aligned databases to predict Néel/Curie temperatures and magnetic ordering.
• Optionally fine-tune a pre-trained CGCNN (transfer learning) on the aligned database.
• Evaluate using MAE, R^2, and CCR; validate with MagNData benchmark.

Experimental results

Research questions

• RQ1 Does aligning experimental magnetic-material data with complete CIF structures improve predictive accuracy for magnetic properties over composition-only approaches?
• RQ2 What is the impact of alignment strictness (composition vs composition+space group) on predictive performance?
• RQ3 Can transfer learning from related formation-energy tasks further enhance magnetic-property predictions when data are limited?
• RQ4 How does the aligned database perform against hand-curated benchmarks like MagNData?

Key findings

• CGCNNs trained on alignment-by-composition+space-group data substantially improve Curie/Néel temperature predictions compared to composition-only or no-alignment baselines.
• Transfer learning from a formation-energy pre-trained CGCNN yields further gains in predictive accuracy, especially with smaller datasets.
• Validation on MagNData shows good agreement between predicted and actual Néel temperatures, with MedAE comparable to training-set MAE.
• The alignment process reduces mismatch between CIF structures and magnetic-property entries, enabling richer graph representations for ML models.
• The approach demonstrates that complete structural information is crucial for accurate magnetic-property prediction in ML models.