Predicting Chemically Accurate Adsorption Energy Using an Interpretable Deep Learning Model Pretrained by GGA Calculation Data

Abstract: Molecular adsorption energy is a critical descriptor for high-throughput screening of heterogeneous catalysts and electrode materials. However, precise experimental adsorption energies are scarce due to the complexity of experiments, while density functional theory (DFT) calculations remain computationally expensive for large-scale material screening. Machine learning models trained on DFT data have emerged as a promising alternative, but face challenges such as functional dependency, limited labeled high-fidelity data, and interpretability issues. Herein, we present DOS Transformer for Adsorption (DOTA), a novel deep learning model that leverages local density of states (LDOS) as input to predict chemically accurate adsorption energies across metallic and intermetallic surfaces. DOTA integrates multi-head self-attention mechanisms with LDOS feature engineering to capture orbital interaction patterns, achieving superior accuracy and transferability. Pretrained on PBE-level DFT data, DOTA can be fine-tuned using minimal high-fidelity experimental or hybrid functional data to predict adsorption energies with chemical accuracy. The model resolves long-standing challenges, such as the "CO puzzle" and outperforms traditional theories, like the d-band center and Fermi softness models. It provides a robust framework for efficient catalyst and electrode screening, bridging the gap between computational and experimental approaches.