BASIL: Bayesian Application for Scientific Iteration and Learning

Technical Summary of BASIL: Bayesian Application for Scientific Iteration and Learning

Problem Statement
Process optimization across scientific and industrial domains traditionally relies on methods like factorial design, simulated annealing, and response surface methodology. While effective in specific contexts, these approaches often struggle with generalization and interpretability. Recently, Bayesian Optimization (BO) has emerged as a powerful alternative, balancing the exploration of experimental design regions with the exploitation of historical data to minimize experimental costs and maximize efficiency. However, a significant barrier to the adoption of BO exists: most existing frameworks (e.g., ProcessOptimizer, BayBE, BOTorch) require users to possess proficiency in Python, command-line interfaces, or machine learning algorithms. This technical hurdle prevents many experimental scientists and technicians from integrating these advanced optimization strategies into their workflows.

Methodology
To address this accessibility gap, the authors introduce BASIL (Bayesian Application for Scientific Iteration and Learning), a user-friendly desktop application designed for process optimization. BASIL functions as a graphical interface that abstracts the complexity of model-based optimization.

• Architecture and Design: BASIL adheres to a "file over apps" philosophy, storing all data and results in standard, non-proprietary JSON and CSV formats to ensure compatibility and transferability. It is built to support Windows, Mac, and Linux operating systems.
• Core Engine: The application leverages the BayBE library to support BO routines, providing an interoperable API.
• Surrogate Modeling: BASIL implements Gaussian Processes (GPs) as the primary surrogate model, utilizing pluggable kernel choices and numerically robust matrix routines to compute posterior mean and variance. It also supports ensemble models.
• Acquisition and Optimization: The system incorporates various acquisition functions (e.g., Expected Improvement, Upper Confidence Bound, logNoisyExpectedHypervolumeImprovement). The optimization process employs a mixture of random restarts and local gradient-based optimization to locate high-value candidates.
• Parameter Handling: The software manages diverse parameter types, including continuous, discrete, and categorical variables. It handles value normalization, encoding (ordinal or one-hot), and decoding to ensure proposed candidates can be translated back to experimental controllers.
• Workflow: Users define a "campaign" by specifying input parameters, boundaries, and optimization objectives (single or multi-objective). The system can ingest legacy data or begin with a random batch for initial space exploration. Based on entered experimental results, the acquisition function suggests the next batch of parameters.
• Visualization and Explanation: BASIL includes tools for 2D and 3D visualization of parameter-outcome relationships (including colormaps) and provides model explanation via SHAP (SHapley Additive exPlanations) values to quantify parameter contributions.

Key Contributions

1. Accessibility: BASIL provides a graphical interface that allows scientists and technicians to perform model-based optimization without writing code or possessing prior knowledge of Python or machine learning.
2. Versatility: It supports both single-objective and multi-objective optimization problems. For multi-objective cases, users can choose to optimize a weighted sum of objectives or identify the Pareto frontier.
3. Comprehensive Feature Set: The application offers predefined surrogate model templates, multiple acquisition functions, and robust handling of mixed variable types (continuous, discrete, categorical).
4. Data Management: It facilitates research data management by tracking all generated and uploaded experimental data, simplifying downstream reporting and analysis.
5. Interoperability: By using standard file formats and supporting external experiment runners (synchronous or asynchronous), BASIL can be integrated with simulated test problems or real laboratory hardware.

Results
The authors demonstrate the functionality of BASIL through a synthetic case study involving the multi-objective optimization of a chemical reaction.

• Setup: The campaign involved five parameters: two continuous (Temperature, pH), two discrete (Catalyst amount, Stirring speed), and one categorical (Catalyst type). The objectives were to maximize Yield and Purity.
• Simulation: The objective functions were defined as weighted sums of quadratic functions for numerical parameters and constant values for the categorical parameter.
• Performance: The study showed that a user could optimize the reaction in only 5 iterations, with a batch size of 10 experiments per iteration.
• Visualization: The results highlighted the software's ability to visualize relationships between parameters and outcomes (e.g., Temperature vs. pH vs. Purity) and to explain model behavior using SHAP beeswarm plots, quantifying how each parameter influenced the objectives.

Significance and Claims
The paper positions BASIL as a critical tool for democratizing advanced process optimization. Its primary significance lies in removing the "adoption hurdle" for experimental researchers who lack computational expertise. By providing an intuitive interface that handles the complexities of Bayesian Optimization, BASIL enables a wider adoption of these methods in scientific discovery and industrial process improvement. The authors claim that the software makes sophisticated optimization accessible to researchers with no computer expertise, thereby bridging the gap between theoretical optimization algorithms and practical experimental workflows. The software is made freely available for academic and non-commercial purposes, with commercial licensing available through the University of North Carolina at Chapel Hill.