Multi-fidelity active learning for cost-aware planning of directed self-assembly simulations

Polymer simulations inevitably involve a trade-off between the accessible spatiotemporal scales of a simulation and its computational cost. We present a cost-aware active learning workflow that couples a dissipative particle dynamics (DPD)  simulation engine with a multi-fidelity Bayesian surrogate model to search for optimal design parameters across a continuous cost-fidelity spectrum. Our oracle is a DPD-based directed self-assembly (DSA) engine in which simulation box size and production run length are tuned independently, spanning from low-cost/low-fidelity to high-cost/high-fidelity evaluations. Guided by a multi-fidelity knowledge gradient acquisition function, we find that strategies that incorporate many low-cost evaluations consistently outperform single-fidelity active learning restricted to high-cost simulations. We further identify cost model design as a critical factor, showing that dynamic rescaling of the cost ratio provides an effective strategy to avoid over-sampling of low-fidelity evaluations. This framework could extend to integrating multiple data sources that differ in cost and fidelity, accelerating the discovery of high-performance materials and manufacturing processes.