Machine Learning-Assisted Performance Prediction and Design of MXene Materials
TL;DR
Here, reported ML-assisted studies on MXene prediction and design are organized by target property and workflow, and ML–DFT screening, graph neural networks, uncertainty quantification, active learning, and interpretability are examined.
Abstract
MXenes are not a single fixed material family, but a broad set of two-dimensional transition-metal carbides and nitrides. Their conductivity, hydrophilic surfaces, layered structure, and adjustable terminations explain why they are frequently studied for electrocatalysis, energy storage, and multifunctional devices. The same flexibility also makes their design difficult. Changing the metal element, carbide/nitride framework, surface group, defect state, interlayer environment, or heterointerface may alter the final property. Exhaustive trial-and-error experiments or one-by-one DFT calculations are therefore inefficient. In this setting, machine learning (ML) combined with density functional theory (DFT) offers a practical way to screen structures, estimate properties, and narrow the candidate space. Here, reported ML-assisted studies on MXene prediction and design are organized by target property and workflow. The electrocatalysis part covers adsorption free energies, overpotentials, catalytic activity, and complex active sites. The energy-storage part discusses capacity, voltage, capacitance, and electronic-structure descriptors. The later sections examine ML–DFT screening, graph neural networks, uncertainty quantification, active learning, and interpretability. The literature still shows several weak points: small curated datasets, inconsistent data standards, simplified treatment of mixed terminations and defects, and limited transfer across MXene systems. More realistic descriptors, standardized datasets, multi-target screening, interpretable models, and tighter experiment-ML–DFT feedback are needed.
