Infrastructure-Aware Materials Discovery: Resource Allocation, Cost, and Carbon in AI-Driven Workflows

Machine-learned interatomic potentials have converted materials-discovery campaigns from tens to hundreds of expensive density-functional calculations into populations of 10^4 to 10^6 short, independent evaluations, relocating the dominant constraint from the electronic-structure solver to the orchestration layer beneath it. We organize AI-driven materials-discovery workloads into four recurring classes—short independent screening tasks, long-running simulations, tightly coupled ab initio calculations, and iterative active-learning loops—and map spot and preemptible computing, checkpointing and workflow-level fault tolerance, autoscaling, cloud bursting, cost-aware Bayesian optimization, and carbon-aware scheduling against this taxonomy. High-throughput screening of metal–organic frameworks serves as a case study arguing that the optimal balance of elastic, stable, and interruptible resources follows from task duration, computational coupling, and recovery cost rather than from generic cloud-optimization heuristics. We further argue that the realized cost of a materials calculation decomposes into scientific, provisioning, recovery, and data-movement terms, of which only the first is a property of the calculation itself. Because the remaining terms vary with market and grid conditions on timescales of minutes, an acquisition function that prices candidates by fidelity alone systematically misprices the campaign it is scheduling. Integrating infrastructure cost and carbon signals into scientific decision-making is identified as a key open problem for autonomous discovery.