Best Practices for Using Genetic Algorithms in Molecular Discovery

Abstract

Genetic algorithms (GAs) are a powerful tool to search large chemical spaces for inverse molecular design. However, GAs have multiple hyperparameters that have not been thoroughly investigated for chemical space searches. In this work, we examine the general effects of a number of hyperparameters, such as population size, elitism rate, selection method, mutation rate, and convergence criteria, on key GA performance metrics. We show that using a self-termination method with a minimum Spearman's rank correlation coefficient of 0.8 between generations maintained for 50 consecutive generations along with a population size of 32, 50% elitism rate, 3- way tournament selection, and a 40% mutation rate provides the best balance of finding the overall champion, maintaining good coverage of elite targets, and improving relative speedup for general use in molecular design GAs.

Screening Efficient Tandem Organic Solar Cells with Machine Learning and Genetic Algorithms

“…It is possible that there are better candidates within the search space; however, they have not been found due to the stochastic nature of the GA. Recent work by our group has shown that tuning the GA hyperparameters, such as mutation rate, selection method, and convergence criteria, can affect the efficiency of searching the chemical space …”

“…Recent work by our group has shown that tuning the GA hyperparameters, such as mutation rate, selection method, and convergence criteria, can affect the efficiency of searching the chemical space. 57 The massive speedups of using GAs to efficiently screen chemical space compared to brute-force high-throughput screening approaches can be seen in Table 1. The largest search space is ca.…”

Using genetic algorithms to discover novel ground-state triplet conjugated polymers

“…We have used the same hyperparameters for the GA from a previous study done in our group as they have shown to be effective for similar molecular systems. 15,16 The GA terminates the Selection-Crossover-Mutation cycle after 40 generations, which we have found to be sufficient in finding the minimal calculated HOMO-LUMO gap (Fig. 2).…”

“…There are ways to mitigate this behavior by tuning the GA's hyperparameters, such as the population size, mutation rate, and elitism rate. 16 However, even with well-tuned hyperparameters, there is still a chance that the GA misses the global optima. While there are other, deterministic algorithms that can find the global optima, they come with a greater computational cost.…”

Metal–Organic Framework Stability in Water and Harsh Environments from Data-Driven Models Trained on the Diverse WS24 Data Set

“…Each GA was run for 20 generations with a population size of 100 and crossover probability, gene-wise mutation probability, and elitism rate of 0.9, 0.4, and 0.5, respectively. 135 The individuals in the initial population were built randomly. For each generation, the fitness values for the population were computed as follows: CIFs for each individual were generated using PORMAKE 0.2.0; 114 features were generated using molSimplify 1.7.3 123,125 and Zeo++ 0.3; 129 the solvent removal stability, 73 2-class water stability, and acid stability models were used to make predictions on the features; and the product of stable class probabilities from the three models was used as the fitness metric, as mentioned previously (eq 1 and SI Figure S18).…”