Source-linked AI summary
Infusing Theory into Deep Learning for Interpretable Reactivity Prediction
Shih-Han Wang, Hemanth Somarajan Pillai, Siwen Wang, Luke E. K. Achenie, Hongliang Xin
TL;DR
Regression-based catalyst design faces ambiguous physical interpretation, while some hand-crafted features require fully optimized clean-site structures, adding computational overhead to reactivity prediction. TinNet combines a convolutional regression module with a d-band-theory chemisorption module to predict adsorption properties of transition-metal sites. TinNet provides detailed physical interpretation of metal-site reactivity inaccessible to regression-based models, with 10-fold cross-validated MAEs of 0.147 eV for *O and 0.116 eV for *N.
Problem
Regression-based catalyst design faces ambiguous physical interpretation, while some hand-crafted features require fully optimized clean-site structures, adding computational overhead to reactivity prediction.
Source: Results · p. 2Method
TinNet combines a convolutional regression module with a d-band-theory chemisorption module to predict adsorption properties of transition-metal sites.
Source: Results · p. 1Results
TinNet provides detailed physical interpretation of metal-site reactivity inaccessible to regression-based models, with 10-fold cross-validated MAEs of 0.147 eV for *O and 0.116 eV for *N.
Source: Results · pp. 5–6Takeaways & Limitations
The paper concludes TinNet can be transferred directly to other descriptor species and nanostructures with different site geometries or electronic complexities.
Source: Results · p. 6Training deployable models with rigorous validation remains expensive, requiring 10^2–10^3 GPU hours.
Source: Results · p. 5
Abstract
from arXiv · showhide
Despite recent advances of data acquisition and algorithms development, machine learning (ML) faces tremendous challenges to being adopted in practical catalyst design, largely due to its limited generalizability and poor explainability. Here, we develop a theory-infused neural network (TinNet) approach that integrates deep learning algorithms with the well-established $d$-band theory of chemisorption for reactivity prediction of transition-metal surfaces. With simple adsorbates (e.g., *OH, *O, and *N) at active site ensembles as representative descriptor species, we demonstrate that the TinNet is on par with purely data-driven ML methods in prediction performance, while being inherently interpretable. Incorporation of scientific knowledge of physical interactions into learning from data sheds further light on the nature of chemical bonding and opens up new avenues for ML discovery of novel motifs with desired catalytic properties.
Results
TinNet combines graph-based learning with d-band chemisorption theory to predict adsorption properties while exposing interpretable bonding quantities. Results cover *OH benchmarks, unseen single-atom alloys, and *O/*N cases across different site settings.
Model benchmark
- As training samples increased, TinNet attained a 0.118 eV MAE with a .022 eV deviation for *OH adsorption energies, outperforming FCNN (0.152±.015 eV) and matching CGCNN (0.114±.025 eV).Source: Results · p. 3
Model validation with single-atom alloys
- On unseen Cu-, Ag-, and Au-host single-atom alloys, TinNet achieved 0.161±.008 eV MAE, beating FCNN (0.193±.026 eV) and CGCNN (0.179±.029 eV), especially for early transition metals.Multi-task FCNN and CGCNN errors remained similar or slightly worsened: 0.198±.039 eV and 0.185±.029 eV, respectively.Source: Figure 3 · p. 4, Results · pp. 4–5
Discussion Model interpretability
- TinNet’s decomposition shows orbital hybridization dominates *OH binding, while Pauli repulsion matters for late transition metals with nearly filled d-states.The d-band center and upper edge were not linearly correlated with *OH adsorption energy (R2: 0.64 and 0.49); βV^2_ad correlated with hybridization energies for 3σ, 1π, and 4σ* orbitals (R2 ∼0.93, ∼0.87, and ∼0.89).Source: Figure 4 · p. 5, Results · p. 5
TinNet models for other adsorbates/facets
- Cross-validated TinNet MAEs were .147 eV for *O atop {111} alloys and .116 eV for *N at hollow sites on {100} Pt-based ternary alloys.The *N dataset comprised 329 surfaces with coadsorbed *OH as a spectator species.Source: Results · p. 5, Figure 5 · p. 6
Conclusion
- TinNet is presented as transferable to other descriptor species and nanostructures with different site geometries or electronic complexity, including compounds with strongly correlated d electrons.The authors frame this as a path toward interpretable discovery of catalytic motifs with desired properties.Source: Results · p. 6