Neuroevolution Potential

Updated 2 February 2026

Neuroevolution is a set of algorithms that optimize neural network weights and topology through evolutionary computation, offering robustness to local minima and hyperparameter stability.
It employs methods such as SNES, NEAT, and quality diversity techniques like MAP-Elites to drive architecture search, skill discovery, and diverse behavior under evolving environments.
Recent advancements integrate surrogate modeling, physics-informed training, and transfer learning to achieve near-DFT accuracy in materials simulation and enhance performance in complex control tasks.

Neuroevolution is an umbrella term for a class of algorithms that train artificial neural networks using evolutionary computation—genetic algorithms, evolution strategies, or related population-based optimizers—either for weights alone ("direct encoding"), or simultaneously for both weights and topology ("indirect encoding" or "architecture search"). Contrary to gradient-based learning, neuroevolution excels in robustness to local minima, hyperparameter stability, architectural adaptation, and skill transfer, making it a competitive paradigm for diverse fields, including control, atomistic simulation, and scientific machine learning.

1. Algorithmic Foundations and Evolutionary Schemes

Neuroevolution proceeds by maintaining a population of candidate networks, encoded as parameter vectors, network graphs, or compositional pattern-producing networks (CPPNs). The central loop applies variation operators—mutation (weight perturbation, topology modification) and crossover—followed by selection based on task-specific fitness functions.

A canonical example is the Separable Natural Evolution Strategy (SNES), which samples parameters $z_k$ from a multivariate Gaussian $\mathcal{N}(m, \sigma^2)$ , evaluates their fitness (often negative validation loss), and updates $m, \sigma$ via natural gradient estimators derived from the population (Fan et al., 2021, Liang et al., 30 Apr 2025). Topology evolution, such as NeuroEvolution of Augmenting Topologies (NEAT), employs genotype encodings with historical marking for gene alignment, speciating networks by compatibility distance and applying complexification via add-node and add-connection mutations (Gajurel et al., 2018, Alcaraz-Herrera et al., 2024).

The evolutionary training loop is highly parallelizable. Frameworks such as EvoJAX exploit hardware acceleration, compiling the ask–evaluate–tell interface via JAX's single-program-multiple-data primitives to achieve near-linear scaling across multiple GPUs/TPUs (Tang et al., 2022).

2. Quality Diversity and Skill Discovery

Historically, reinforcement learning was dominant for neural policy training, but deep RL’s sensitivity to hyperparameters and failure modes under environmental perturbations spurred interest in Quality Diversity (QD) neuroevolution (Chalumeau et al., 2022). QD algorithms, including MAP-Elites and policy-gradient hybrids (PGA-MAP-Elites, CMA-ME), maintain an explicit archive of policies distributed over discretized behavioral spaces $\mathbb{M}$ . Each cell contains the “elite” with the highest fitness for its descriptor, guaranteeing coverage and diversity.

Empirical comparisons show that vectorized QD neuroevolution reliably discovers broader repertoires of skills and adapts more robustly to domain shifts—for example, maintaining higher returns under actuator damage or gravity changes, where RL skill-mixes collapse. MAP-Elites hyperparameter robustness is 2–5× greater than RL methods, and evolution-driven planners (meta-PPO) exploiting QD skill libraries outperform RL both for downstream composition and extreme perturbations (Chalumeau et al., 2022).

3. Representation Learning, Architectural Search, and Surrogate Methods

Neuroevolution is not limited to fixed-topology weight search. Linear Genetic Programming (LGP) forms a genotype of register-based instructions, enabling automatic assembly of variable-depth, branching DNN architectures (Stapleton et al., 2024). Combined with surrogate modeling—such as Kriging Partial Least Squares (KPLS) regression on intermediate features—evolution can be accelerated by up to 25%, as noisy full evaluations are replaced with reliable, computationally cheap fitness estimates. Surrogate-assisted neuroevolution matches fully expensive methods in test accuracy, outperforming baselines such as VGG-16 and further advancing practical neuroevolution-directed architecture search.

4. Neuroevolution Potentials for Atomistic and Materials Simulation

In atomistic simulation, the Neuroevolution Potential (NEP) architecture constructs interatomic potentials by decomposing the total energy into local, symmetry-invariant neural network site energies $U_i(\mathbf{q}_i)$ . Descriptors $\mathbf{q}_i$ leverage radial and angular basis expansions (often Chebyshev polynomials and Legendre/spherical harmonics) over local atomic environments, with scalable implementations yielding $>10^7$ atom-step/s per GPU (Fan et al., 2021, Cao et al., 19 May 2025, Liang et al., 30 Apr 2025).

The neuroevolutionary optimizer jointly tunes network weights and descriptor coefficients via SNES, achieving near-DFT accuracy (energy RMSE $<10$ meV/atom, force RMSE $<100$ meV/Å) and empirical-potential speeds. Descriptor learning, extending trainable type-pair and channel-specific weights, further reduces regression error and enables accurate extension to multi-component alloys, ceramics, and organics (Fan, 2021).

Foundation models such as NEP89 are trained on curated, descriptor-space-subsampled datasets spanning 89 elements, delivering accurate simulation for metals, alloys, organics, and biomolecules. The architecture accommodates fine-tuning for specialist domains with $<200$ DFT-labeled structures. NEP-based MD matches experimental and high-level simulations in structural, mechanical, thermal, and spectroscopic properties, at scales exceeding 10 million atoms (Liang et al., 30 Apr 2025).

5. Extension to Physical Properties, Spectroscopy, and Electrostatics

Advances in NEP frameworks generalize to tensorial observables (dipoles, polarizabilities, susceptibilities) via TNEP, enabling rapid, on-the-fly prediction of IR and Raman spectra in liquids and solids (Xu et al., 2023). Extension to dynamic, environment-dependent charges (qNEP) leverages neural prediction of partial and Born effective charges from local descriptors, permitting direct computation of polarization and dielectric tensors within Ewald or PPPM summation schemes (Fan et al., 26 Jan 2026).

NEP-D3 combines NEP short-range chemistry with long-range Grimme D3 dispersion within a unified, GPU-parallel architecture. This hybrid achieves accurate modeling of nonbonded interactions, reproducing binding and sliding energies in van-der-Waals systems and correcting thermal conductivities in MOFs with $<$ 0.05 eV/Å force RMSE against reference DFT-D3 (Ying et al., 2023).

6. Transfer Learning, Multi-Agent Control, and Automated Design

Neuroevolution's modularity and genetic diversity foster transfer learning and adaptation. Empirical benchmarks demonstrate superior skill transfer across task curricula and morphologies compared to RL baselines (Nisioti et al., 28 May 2025). In soft robotics, neuroevolutionary search over actuator morphologies (NEAT, HyperNEAT, AFPO) yields robust, non-intuitive catheter-tip designs, outperforming hand-designed experts in displacement, compactness, and robustness under control uncertainty (Alcaraz-Herrera et al., 2024).

In game-theoretic micro (RTS), topology evolution produces networks capable of emergent tactics (kiting), generalizing over unseen agent configurations—demonstrating spontaneous behavioral synthesis at compact representation scales previously unattainable via manual scripting or RL-only approaches (Gajurel et al., 2018).

7. Neuroevolution in Scientific ML and Physics Compliance

Physics-informed neural networks (PINNs), which embed PDE residuals directly into their loss, pose significant optimization challenges due to rugged, multimodal loss landscapes and deceptive local minima. Population-based neuroevolutionary algorithms (CMA-ES, xNES+NAG) outperform gradient descent by escaping spurious optima, delivering lower physics-residuals and higher compliance in complex governed domains. Hardware-accelerated, vectorized evolutionary frameworks (e.g., EvoJAX) render large-population, physics-aware training practical, opening PINNs to high-dimensional industrial and multiphysics deployments (Yong et al., 2022).

In summary, neuroevolution unifies global search, robustness to nonconvex loss geometries, scalable vectorized implementation, and flexible representation learning. It underpins advances in control, skill transfer, automated design, atomistic simulation, and scientific discovery, with empirical performance routinely matching or surpassing state-of-the-art gradient- and RL-based methods in adaptation, diversity, accuracy, and computational efficiency (Chalumeau et al., 2022, Tang et al., 2022, Liang et al., 30 Apr 2025, Fan et al., 2021).