Supported Algorithms

CyRxnOpt supports a variety of state-of-the-art optimization algorithms, ranging from traditional methods to modern machine learning-driven approaches. These algorithms enable efficient exploration of reaction condition landscapes, local and global search strategies, and single- or multi-objective decision-making.

Below are sections briefly describing each supported algorithm, along with the CyRxnOpt ID associated with each method in CyRxnOpt.

AMLRO

CyRxnOpt ID: amlro

AMLRO (Active Machine Learning Reaction Optimizer) is our in-house active learning framework designed for reaction space optimization. [amlro] It integrates regression models with iterative candidate selection. AMLRO is capable of handling continuous and categorical features, dynamic objectives with directions (minimization or maximization), and user feedback loops.

  • Status: Source code is currently under development. Public release is planned, and the project will be hosted on GitHub (expected link: TBA).

References

[amlro]

Manuscript in preparation.

EDBO+

CyRxnOpt ID: edbop

EDBO+ [garridotorres2022] is a Bayesian optimization framework specifically designed for chemical reaction optimization based on the original EDBO (Experimental Design via Bayesian Optimization) algorithm [shields2021] from the Doyle group. It leverages Gaussian process models with acquisition functions (e.g., Expected Improvement, qEI, and EHVI for multi-objective settings) to identify new experiments that are most informative. EDBO+ has been successfully applied to global, multi-objective optimization of reaction yields, selectivity, and sustainability metrics.

References

[shields2021]

Shields, B. J.; Stevens, J.; Li, J.; Parasram, M.; Damani, F.; Alvarado, J. I. M.; Janey, J. M.; Adams, R. P.; Doyle, A. G. Bayesian Reaction Optimization as a Tool for Chemical Synthesis. Nature 2021, 590 (7844), 89–96. DOI: 10.1038/s41586-021-03213-y.

[garridotorres2022]

Garrido Torres, J. A.; Lau, S. H.; Anchuri, P.; Stevens, J. M.; Tabora, J. E.; Li, J.; Borovika, A.; Adams, R. P.; Doyle, A. G. A Multi-Objective Active Learning Platform and Web App for Reaction Optimization. J. Am. Chem. Soc. 2022, 144 (43), 19999–20007. DOI: 10.1021/jacs.2c08592.

Nelder-Mead Simplex

CyRxnOpt ID: nmsimplex

CyRxnOpt uses Scipy’s Python implementation of the Nelder-Mead Simpex method (NMSIM). [nelder1965] NMSIM is a derivative-free, local optimization algorithm widely used for smooth, low-dimensional problems. It iteratively refines a simplex (a polytope with n+1 vertices in n-dimensional space) through reflection, expansion, contraction, and shrink steps to approach a local minimum. While efficient for surfaces or regions with a single optimum, noisy surfaces or those with many local optima may converge to a local optimum rather than the desired global solution.

References

[nelder1965]

Nelder, J. A.; Mead, R. A Simplex Method for Function Minimization. The Computer Journal 1965, 7 (4), 308–313. DOI: 10.1093/comjnl/7.4.308.

SQSnobFit

CyRxnOpt ID: sqsnobfit

SQSnobFit is scikit-quant’s Python implementation of Huyer and Neumaier’s original SNOBFIT (Stable Noisy Optimization by Branch and Fit) algorithm. [huyer2008] SNOBFIT is a global optimization algorithm particularly well-suited for expensive, noisy black-box functions. It combines local quadratic model fitting with global space-partitioning to balance exploration and exploitation.

References

[huyer2008]

Huyer, W.; Neumaier, A. SNOBFIT – Stable Noisy Optimization by Branch and Fit. ACM Transactions on Mathematical Software 2008, 35 (2), 1–25. DOI: 10.1145/1377612.1377613.