Optimizing Methane Uptake on N-O Functionalized Graphene via DFT, Machine Learning, and Uniform Manifold Approximation and Projection (UMAP) Techniques
 Techniques/hero.png)
This work develops a hybrid DFT-machine learning pipeline to model and rediscover methane adsorption on functionalized carbon materials. The chemical systems considered are graphene-based surfaces decorated with nitrogen and oxygen functionalities. From a non-technical standpoint, these surface groups slightly modify how strongly methane molecules attach to the material. The goal was to understand and predict which surface configurations optimize methane uptake.
On the computational side, adsorption energies were first computed using DFT (PW91, DMol³), providing a structured dataset of methane–surface interactions across multiple functional groups and geometries. These data were then used to train and evaluate multiple ML models, including neural networks, regression models, and classifiers. The models achieved strong predictive performance for adsorption energy (R² ≈ 0.99), indicating that the learned representations effectively capture the underlying structure–property relationships.
A key contribution lies in the integration of:
Multi-model learning to compare predictive stability across algorithms.
Sensitivity analysis to quantify the influence of structural and chemical descriptors (e.g., deformation energy, functionality type, optimized geometry). The resulting energy sensitivities were small (on the order of ±0.02–0.03 eV), indicating physically consistent perturbative behavior.
ML-assisted rediscovery, where the trained models were used to re-identify optimal adsorption configurations from the design space.
UMAP-based embedding as a complementary screening tool. The low-dimensional manifold representation provides geometric organization of adsorption configurations and supports candidate selection beyond direct regression outputs.
The study demonstrates that combining DFT-generated datasets with structured ML workflows enables high-fidelity prediction and efficient screening of adsorption materials. Methodologically, the main contribution is the synergistic integration of predictive modeling, interpretability via sensitivity analysis, and manifold learning for materials rediscovery.