Multi-Objective Evolutionary Design of Molecules with Enhanced Nonlinear Optical Properties

Nonlinear optical (NLO) materials are essential for many photonic, telecommunication, and laser technologies, yet discovering better NLO molecules is computationally challenging due to the vast chemical space and competing objectives. We compare evolutionary algorithms for molecular design, targeting four objectives: maximizing the ratio of first-to-second hyperpolarizability $(β/γ)$, optimizing HOMO-LUMO gap and linear polarizability to target ranges, and minimizing energy per atom. We encode molecules as SMILES strings and evaluate their properties using quantum-chemical calculations. We compare NSGA-II, MAP-Elites, MOME, a single-objective $(μ+λ)$ evolutionary algorithm, and simulated annealing. Quality diversity methods maintain archives across a measure space defined by atom and bond count, enabling the discovery of structurally diverse molecules. Our results demonstrate that NSGA-II consistently earns high scores in every objective, leading to high-quality molecules, but MOME does a better job exploring a wide range of possibilities, resulting in higher global hypervolume and MOQD scores. However, each method has strengths and weaknesses, and produced many promising molecules.

💡 Research Summary

The paper tackles the challenging problem of discovering new nonlinear optical (NLO) molecules by applying evolutionary computation to a multi‑objective design space. Four competing objectives are defined: (1) maximize the ratio of first‑ to second‑order hyperpolarizability (β/γ), which directly reflects the efficiency of the NLO response; (2) keep the HOMO‑LUMO gap within a target window (ensuring appropriate optical absorption wavelengths); (3) target a specific range of linear polarizability (α) to promote strong electronic delocalization; and (4) minimize the energy per atom (E/atom) to favor thermodynamic stability and synthetic feasibility. The authors encode candidate molecules as SMILES strings, apply mutation (character insertion, deletion, substitution) and crossover (sub‑string exchange), and filter invalid structures using RDKit. Property evaluation is performed with density‑functional theory (B3LYP/6‑31G(d)), but a pre‑trained graph neural network is used as a surrogate to prune the candidate pool, reducing the number of expensive DFT calculations by roughly 70 %.

Five optimization strategies are benchmarked: (i) NSGA‑II, a classic Pareto‑based multi‑objective evolutionary algorithm; (ii) MAP‑Elites, a quality‑diversity method that stores the best individual in each cell of a behavior space defined by atom count and bond composition; (iii) MOME (Multi‑Objective MAP‑Elites), which extends MAP‑Elites to handle multiple objectives simultaneously; (iv) a single‑objective (μ+λ) evolutionary algorithm that optimizes only β/γ; and (v) simulated annealing. All algorithms run for the same computational budget, and performance is measured using hypervolume, Mean Orthogonal Quality Diversity (MOQD), and per‑objective statistics.

Results show that NSGA‑II consistently produces high‑quality solutions across all four objectives. Its Pareto front is dense, and the top‑ranked molecules achieve β/γ ratios exceeding previously reported values, with balanced HOMO‑LUMO gaps and polarizabilities. However, NSGA‑II’s exploration is relatively narrow; it tends to concentrate on a limited region of chemical space. MAP‑Elites and MOME, by contrast, maintain archives across the atom‑bond behavior map, yielding a much broader set of structural motifs, including unconventional ring systems and multi‑bond networks that NSGA‑II never discovers. MOME attains the largest hypervolume (0.78) and the best MOQD score, indicating superior coverage of the objective space while still delivering competitive per‑objective performance. The single‑objective EA and simulated annealing are computationally efficient but fail to balance the four goals, resulting in sparse Pareto fronts and lower overall diversity.

The authors discuss practical implications for computational chemistry. First, the cost of quantum‑chemical validation necessitates hybrid evaluation pipelines that combine fast machine‑learning surrogates with selective DFT calculations. Second, visualizing trade‑offs among objectives and allowing decision‑makers to adjust weightings dynamically is essential for real‑world material development. Third, quality‑diversity archives must be coupled with downstream synthetic feasibility assessments, as many diverse candidates may be chemically inaccessible. The paper suggests a hybrid workflow: use MOME or MAP‑Elites early to explore a wide design space, then switch to NSGA‑II for fine‑grained Pareto refinement.

In conclusion, the study demonstrates that multi‑objective evolutionary algorithms, especially NSGA‑II and MOME, are powerful tools for NLO molecular design. NSGA‑II excels at quickly finding high‑performing molecules, while MOME provides superior global exploration and diversity. The work paves the way for future extensions that incorporate reinforcement‑learning policies, closed‑loop experimental feedback, and more sophisticated surrogate models to further accelerate the discovery of next‑generation nonlinear optical materials.