Abstract
Machine learning interatomic potentials (MLIPs) have become a workhorse ofmodern atomistic simulations, and recently published universal MLIPs,pre-trained on large datasets, have demonstrated remarkable accuracy andgeneralizability. However, the computational cost of MLIPs limits theirapplicability to chemically disordered systems requiring large simulation cellsor to sample-intensive statistical methods. Here, we report the use ofcontinuous and differentiable alchemical degrees of freedom in atomisticmaterials simulations, exploiting the fact that graph neural network MLIPsrepresent discrete elements as real-valued tensors. The proposed methodintroduces alchemical atoms with corresponding weights into the input graph,alongside modifications to the message-passing and readout mechanisms of MLIPs,and allows smooth interpolation between the compositional states of materials.The end-to-end differentiability of MLIPs enables efficient calculation of thegradient of energy with respect to the compositional weights. Leveraging thesegradients, we propose methodologies for optimizing the composition of solidsolutions towards target macroscopic properties and conducting alchemical freeenergy simulations to quantify the free energy of vacancy formation andcomposition changes. The approach offers an avenue for extending thecapabilities of universal MLIPs in the modeling of compositional disorder andcharacterizing the phase stabilities of complex materials systems.