One Set to Rule Them All: How to Obtain General Chemical Conditions via Bayesian Optimization over Curried Functions

TL;DR

The paper tackles the problem of finding reaction conditions that generalize across multiple substrates by framing optimization as a search over curried functions $f({\mathbf{x}}; {\mathbf{w}})$ with a generality aggregation $\phi$, observed only partially per experiment. It extends Bayesian optimization to handle a discrete task set ${\mathcal{W}}$ and a continuous input space ${\mathcal{X}}$, via currying and a surrogate model $g_k({\mathbf{x}}, {\mathbf{w}})$, and introduces sequential and joint acquisition strategies for selecting $(\mathbf{x}, \mathbf{w})$ at each step. The authors benchmark multiple generality-oriented BO strategies on four real-world chemical reaction datasets, augmented to better reflect practical experiment spaces. They find that increasing the number of considered substrates improves transfer to unseen tasks, and that a simple sequential acquisition strategy (optimize ${\mathbf{x}}$ first, then ${\mathbf{w}}$ with one-step lookahead) often matches or exceeds more complex policies, with exploration of ${\mathcal{X}}$ being the key factor. The work provides CurryBO, an open-source extension to BoTorch, and emphasizes the need for more realistic benchmarks to advance generality-oriented optimization in chemistry and beyond.

Abstract

General parameters are highly desirable in the natural sciences - e.g., chemical reaction conditions that enable high yields across a range of related transformations. This has a significant practical impact since those general parameters can be transferred to related tasks without the need for laborious and time-intensive re-optimization. While Bayesian optimization (BO) is widely applied to find optimal parameter sets for specific tasks, it has remained underused in experiment planning towards such general optima. In this work, we consider the real-world problem of condition optimization for chemical reactions to study how performing generality-oriented BO can accelerate the identification of general optima, and whether these optima also translate to unseen examples. This is achieved through a careful formulation of the problem as an optimization over curried functions, as well as systematic evaluations of generality-oriented strategies for optimization tasks on real-world experimental data. We find that for generality-oriented optimization, simple myopic optimization strategies that decouple parameter and task selection perform comparably to more complex ones, and that effective optimization is merely determined by an effective exploration of both parameter and task space.

Figures (46)

• Figure 1: Left: While conditions can be optimized to maximize the reaction outcome for only one substrate (red), generality-optimized conditions provide a satisfactory reaction outcome for multiple substrates. Right: Optimization loop for generality-oriented optimization under partial monitoring.
• Figure 2: A conceptual overview of the generality-oriented optimization problem. Left: The function values across the joint space $\mathcal{X} \times \mathcal{W}$. Right: Mean aggregation applied to the function family $f({\mathbf{x}}; {\mathbf{w}})$, that is obtained via currying of the joint space $\mathcal{X} \times \mathcal{W}$. The quantity $\phi(f({\mathbf{x}}; {\mathbf{w}}), {\mathcal{W}})$ constitutes the partially observable objective function, of which $\hat{{\mathbf{x}}} = \mathop{\mathrm{arg\,max}}\limits_{{\mathbf{x}} \in {\mathcal{X}}} \ \phi({\mathbf{x}})$ is the optimum to be identified.
• Figure 3: Normalized test-set generality score as determined by exhaustive grid search for the four benchmarks on the original (left) and augmented (right) problems for the mean aggregation. Average and standard error are taken from thirty different train/test substrates splits.
• Figure 4: Optimization trajectories of different algorithms for generality-oriented optimization previously reported in the chemical domain. The trajectories are averaged over all four augmented benchmark problems. Note that the Bandit algorithm is incompatible with the threshold aggregation function.
• Figure 5: Optimization trajectories using sequential acquisition strategies. The top row shows the variation of $\alpha_x$, while the bottom row shows the variation of $\alpha_w$. Trajectories are averaged over all four augmented benchmark problems. In general, more complex two-step lookahead acquisition strategies outperform more simple one-step lookahead strategies. While more explorative $\alpha_x$ perform better, the choice of $\alpha_w$ does not significantly influence the optimization performance for one-step lookahead strategies.