Table 4 Transfer learning performance comparison

From: LLM-Prop: predicting the properties of crystalline materials using large language models

Model

Volume to Band gap (eV)

Band gap to Volume(˚A³/cell)

Validation set ↓

Test set ↓

Validation set ↓

Test set ↓

Structure based

CGCNN-transfer

0.300

0.295

187.473

182.997

MEGNet-transfer

1.461

1.472

190.013

195.664

ALIGNN-transfer

0.321

0.322

137.471

136.164

Text based

MatBERT-zero-shot

1.175

1.191

412.891

422.487

MatBERT-transfer

0.259

0.266

50.530

54.289

LLM-Prop-zero-shot

1.288

1.103

482.922

485.583

LLM-Prop-transfer

0.236

0.244

47.837

50.753

  1. We directly test models trained on band gap for volume prediction (here denoted with a postfix ”zero-shot”) or further fine-tune them on volume prediction (denoted with a postfix ”transfer”) and vice versa.
Back to article page