Adaptive Computing for Scale-up Problems

TL;DR

This paper presents Adaptive Computing (AC), an outer-loop framework that de-risks scale-up challenges by combining multi-fidelity surrogates, uncertainty management, and automated orchestration of heterogeneous computing and experimental resources within a budget-constrained workflow. It formulates scale-up problems as goal-driven data acquisition tasks, leveraging bridging functions and domain-informed priors to quantify trust and guide adaptive sampling. The approach is demonstrated across diverse renewable-energy applications—biofuels reactor design, autonomous laboratory synthesis, multi-solver coupling for perovskite growth, and building-load control—highlighting improved decision quality under limited budgets and varying resource constraints. The work advances practical scale-up decision making by enabling efficient, resource-aware, multi-fidelity experimentation and computation with online learning.

Abstract

Adaptive Computing is an application-agnostic outer loop framework to strategically deploy simulations and experiments to guide decision making for scale-up analysis. Resources are allocated over successive batches, which makes the allocation adaptive to some objective such as optimization or model training. The framework enables the characterization and management of uncertainties associated with predictive models of complex systems when scale-up questions lead to significant model extrapolation. A key advancement of this framework is its integration of multi-fidelity surrogate modeling, uncertainty management, and automated orchestration of various computing and experimentation resources into a single integrated software package. This enables efficient multi-fidelity modeling across multiple computing resources by incorporating real-world constraints such as relative queue times and throughput on individual machines into the multi-fidelity sampling decision. We discuss applications of this framework to problems in the renewable energy space, including biofuels production, material synthesis, perovskite crystal growth, and building electrical loads.

Figures (5)

• Figure 1: The software drives the scheduling of application-specific simulations. A surrogate model informs an acquisition function, which selects simulation cases. The hardware scheduler manages their execution on computational resources.
• Figure 2: Engineering design optimization for reactor. The contour plot represents a Gaussian process model for glucan concentration trained on Vebio simulations indicated by the symbols. The black x's are three reactor designs with random samples using , and the colored circles are seven sequential samples using with the criteria for maximizing glucan concentration.
• Figure 3: The multi-fidelity experiment/simulation feedback loop for process control design. The experimental feedback loop iteratively grows and etches a thin film using suggested growth conditions from the experimental driver, storing in-situ data products. In parallel, the simulation feedback loop evaluates comparable molecular dynamic simulations and creates intermediate synthetic data products. The AC driver learns a model fidelity correction between the experiment and simulation and facilitates updating corresponding ML models.
• Figure 4: A multi-species compressible flow impinges on a substrate (left). The perovskite crystal growth on the substrate is captured with a solver for adsorption, desorption, chemical reactions, and surface and bulk diffusion. The computed flux of species from the gas phase is passed back to the compressible flow code as a boundary condition for the next timestep. The driver orchestrates the flow of information and the evaluation of the microscale model (either through a or a high-fidelity evaluation).
• Figure 5: The multi-fidelity building modelling and control methodology. The driver (GP surrogate model) is used to adaptiveley update the low-fidelity linear control model based on the current conditions. In parallel, the plant model uses a second driver to switch between models with various fidelities to balance accuracy and computational burden. The high fidelity EnergyPlus model updates these models with new data.