Abstract
Scientific experiments in chemistry and materials science often involve multiple competing goals with strict priorities, such as maximizing reaction yield before minimizing reagent cost. Traditional Multi-Objective Optimization (MOO) methods often waste experimental budget exploring irrelevant trade-offs or struggle to model complex, artificially combined objective functions. This paper introduces BoTier, an auto-differentiable, hierarchical composite objective for Bayesian Optimization. By modeling objectives independently and combining them through a tiered scalarization function, BoTier significantly accelerates the discovery of optimal experimental conditions compared to traditional black-box or Pareto-front methods.
Introduction
In fields like chemistry and materials science, researchers rarely optimize for a single metric. A new drug, for instance, must demonstrate high activity, low toxicity, and favorable bioavailability. These objectives are often in conflict, and evaluating them requires expensive, time-consuming experiments. Furthermore, these objectives typically follow a strict hierarchy: a researcher might prioritize achieving a target product yield first, and only then focus on minimizing catalyst cost or reaction temperature.
Existing Multi-Objective Optimization (MOO) algorithms face two primary challenges:
- Inefficient Exploration: Pareto-front mapping methods (e.g., EHVI) attempt to find all optimal trade-offs. When priorities are already known, this wastes resources on trade-offs that are not of interest.
- Modeling Complexity: Black-box scalarization methods (e.g., Chimera) combine objectives into a single score before model training. This forces the surrogate model to learn a highly complex, non-physical function. Additionally, many hierarchical scores are not auto-differentiable, limiting their use with modern gradient-based optimization frameworks.
Methods: The BoTier Approach
BoTier addresses these issues by introducing a hierarchical scalarization function designed as a composite objective. Unlike "black-box" approaches that train one model on a combined score, BoTier trains separate probabilistic surrogate models (Gaussian Processes) for each individual objective. The hierarchical scoring logic is then applied to the outputs of these models.
Crucially, BoTier utilizes continuously differentiable approximations for mathematical functions like min and the Heaviside step function. This ensures the entire pipeline is auto-differentiable within the PyTorch/BoTorch ecosystem, enabling efficient Monte-Carlo integration to identify the optimal next experiment.
Experimental Evaluation
The authors evaluated BoTier across both synthetic mathematical benchmarks and emulated real-world chemistry problems:
- Analytical Surfaces: Standard MOO benchmarks (BNH, DH4, DTLZ5, ZDT1) augmented with input-dependent objectives to simulate constraints like reagent cost.
- Emulated Chemistry Problems: Four datasets representing autonomous "self-driving lab" scenarios: Suzuki–Miyaura cross-coupling, enzymatic alkoxylation, silver nanoparticle synthesis, and amine monoalkylation.
Key Insights
- Composite vs. Black-Box: Modeling objectives independently (the composite approach) consistently outperformed black-box modeling. It is significantly easier for Bayesian Optimization to model multiple simple distributions than one complex, artificially combined distribution.
- Efficiency through Hierarchy: Encoding known priorities (e.g., "yield > 80% before considering cost") directly into the objective function guides the search toward relevant regions of the design space, saving substantial experimental budget.
- Model Simplicity: The study found that independent single-task Gaussian Processes generally performed as well as or better than more complex multi-output models that attempt to capture cross-correlations between objectives.
Results
BoTier demonstrated superior performance across all benchmarks. In synthetic tests, it required fewer experiments to satisfy hierarchical objectives than Chimera or penalty-based baselines.
In the Suzuki-Miyaura coupling case study—optimizing yield, then cost, then temperature—BoTier was the only algorithm to successfully identify conditions meeting all three thresholds within a 50-experiment budget. In contrast, Pareto-mapping (EHVI) and black-box (Chimera) baselines frequently failed to find satisfactory conditions within the same budget.
Limitations and Future Work
A current limitation of BoTier is the requirement for predefined "satisfaction thresholds" for each objective, which may not be known during early-stage exploratory research. Additionally, the differentiable approximations depend on a user-defined smoothness parameter.
Future work will focus on deploying BoTier in physical "self-driving laboratories" (SDLs). The independent modeling of objectives in BoTier also makes it an ideal candidate for incorporating physics-based priors into individual surrogate models, potentially further increasing sample efficiency in complex material and reaction optimization.
Summary updated