Adaptive Experimental Design to Accelerate Scientific Discovery and Engineering Design

Aryan DeshwalAAAI 2025 · DOI

Introduction

In contemporary scientific and engineering workflows, the path from hypothesis to discovery is often obstructed by the sheer cost and complexity of experimentation. Whether synthesizing a novel chemical compound, optimizing a 3D-printed organ model, or configuring a heterogeneous manycore processor, each evaluation can require days of laboratory work or weeks of high-fidelity simulation. This paper explores a body of research focused on Adaptive Experimental Design (AED)—a framework that leverages Bayesian Optimization to navigate high-dimensional search spaces efficiently, identifying optimal designs with minimal resource expenditure.

The Challenge of High-Dimensional Design

The difficulty in modern design optimization stems from two primary factors. First, the search spaces are frequently high-dimensional and combinatorial, involving complex binary or categorical variables that do not map easily to traditional optimization techniques. Second, the "oracle" (the experiment or simulation) is a scarce resource. We cannot afford exhaustive searches; instead, we must treat every experiment as a strategic investment to maximize information gain.

Methodological Framework

The author proposes a dual-strategy approach to address these challenges: advanced probabilistic modeling and information-theoretic decision-making.

1. Probabilistic Modeling for Combinatorial Structures

To model complex design spaces, the research introduces Dictionary Embeddings. This novel Gaussian Process (GP) model is specifically designed for high-dimensional binary and categorical spaces, utilizing a "dictionary" to map discrete structures into a continuous latent space where optimization becomes more tractable.

Furthermore, the LADDER framework bridges the gap between data-driven deep learning and domain expertise. By combining the latent spaces of deep generative models (such as Variational Autoencoders) with structured kernels designed by experts, LADDER allows the optimization process to benefit from both large-scale data patterns and fundamental physical or chemical principles.

2. Information-Theoretic Decision-Making

Beyond modeling, the research focuses on how to choose the next experiment. The Output Space Entropy (OSE) framework shifts the objective from finding a single "best" point to maximizing the information gain regarding the entire optimal Pareto front. This is particularly crucial in multi-objective optimization, where researchers must balance competing goals like material strength versus weight.

Additionally, the work incorporates Multi-Fidelity Optimization. These policies recognize that experimental data comes at different price points—low-fidelity simulations are cheap but noisy, while high-fidelity lab tests are accurate but expensive. The proposed policies optimize the "information gain per unit resource cost," ensuring that expensive resources are only deployed when they provide the most value.

Theoretical Insights: The Compressed Sensing Connection

A significant contribution of this work is the theoretical bridge built between Gaussian Process modeling and compressed sensing. The author demonstrates that "regret bounds"—the measure of performance loss relative to an omniscient optimizer—are fundamentally linked to the orthogonality of the dictionary used in the embeddings. This insight provides a rigorous mathematical foundation for why certain embedding strategies outperform others in high-dimensional search.

Empirical Validation and Results

The methodologies were rigorously tested across several high-stakes domains:

  • Materials Science: The AED approach successfully identified nanoporous materials for methane storage and noble gas separation by evaluating only a tiny fraction of the potential candidate pool.
  • Additive Manufacturing: In 3D printing, the AI-driven search directly improved the physical quality, specifically the precision and porosity, of biological organ models.
  • Hardware Design: The framework enabled efficient design space exploration for 3D heterogeneous manycore systems, a task traditionally plagued by astronomical configuration counts.
  • AutoML: The information-theoretic policies set new state-of-the-art benchmarks for hyperparameter tuning, outperforming existing multi-fidelity optimization methods.

Future Directions

While the current frameworks offer significant acceleration, the author identifies several frontiers for future research. One primary challenge is the presence of black-box constraints—limitations that are only revealed during the experiment itself (e.g., a material that is theoretically optimal but physically impossible to synthesize).

The next evolution of this work lies in Grey-Box Optimization, which seeks to move beyond treating experiments as opaque boxes. By leveraging side information or partial internal data from simulations, these models could further reduce the number of trials needed to reach a breakthrough. Finally, refining cost-awareness remains a priority, as the relative costs of different experimental fidelities can fluctuate dynamically in real-world laboratory settings.

Conclusion

Adaptive Experimental Design represents a paradigm shift in how we approach discovery. By treating the search for new materials and engineering designs as an information-gathering problem, the methodologies presented here provide a robust roadmap for finding "needles in haystacks" across the scientific landscape.

Summary updated

BibTeX
@inproceedings{deshwal2025adaptive,
  author = {Aryan Deshwal},
  title = {{Adaptive Experimental Design to Accelerate Scientific Discovery and Engineering Design}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35104},
}
Figure