💻 computer science

Evolutionary Extreme Learning Machine of ab-initio Energy Landscapes for Crystal Structure Prediction using Manta Ray Optimization with Levy Flight

This paper proposes an Evolutionary Extreme Learning Machine optimized by Manta Ray Foraging with Levy Flight (EELM-MRFO-LF) to predict formation energies in binary crystal systems, utilizing Levy Flight to enhance population diversity and avoid local optima during the training process.

Original authors: Adrian Rubio-Solis

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adrian Rubio-Solis

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding the "Perfect" Crystal

Imagine you are a chef trying to invent a new recipe. You have two ingredients: Lithium (Li) and Germanium (Ge). You want to mix them in every possible way to find the combination that tastes the best (or in this scientific case, has the lowest energy and is most stable).

In the real world, figuring out the perfect mix usually requires a supercomputer to simulate the physics of atoms. This process, called Density Functional Theory (DFT), is like trying to bake a cake by calculating the exact molecular movement of every egg and grain of flour. It's incredibly accurate, but it takes a long time and costs a lot of computing power.

This paper proposes a shortcut: a "smart guesser" (a machine learning model) that can predict the best mix almost instantly, without needing to run the heavy physics simulation every time.

The Problem: The "Random Guess" Trap

The authors use a type of AI called an Extreme Learning Machine (ELM). Think of an ELM as a very fast student who learns by guessing.

How it works: It randomly picks a bunch of "rules" (weights) to connect the ingredients to the result.
The Catch: Because it guesses randomly, it sometimes gets stuck in a "local trap." Imagine a hiker looking for the lowest point in a valley. If they only look at the small dip right in front of them, they might think they've found the bottom, even though there is a much deeper valley just over the next hill. The standard ELM often stops too early, thinking it found the best answer when it hasn't.

The Solution: The Manta Ray with a Superpower

To fix this, the authors combined the ELM with a search algorithm inspired by nature, called Manta Ray Foraging Optimization (MRFO).

The Analogy:
Imagine a school of Manta Rays searching for a massive patch of plankton (food) in the ocean. They use three clever strategies:

Chain Foraging: They line up head-to-tail. If the one in front finds food, the one behind follows.
Cyclone Foraging: They swim in a spiral toward the food, getting closer and closer.
Somersault Foraging: They flip over the food source to explore the area around it.

The Upgrade (Lévy Flight):
The authors realized that sometimes the Manta Rays get stuck in a small patch of food and miss the giant buffet nearby. To fix this, they added Lévy Flight.

Think of Lévy Flight as giving the Manta Rays a magic compass that occasionally makes them take giant, random leaps. Instead of just swimming slowly in a circle, the compass might suddenly tell a ray to jump 100 miles in a random direction. This ensures they don't get stuck in one spot; they can explore the whole ocean to make sure they find the absolute best food source, not just a good one.

How They Tested It

The team trained this "Manta Ray AI" on a dataset of 14,000 different Lithium-Germanium crystal structures.

The Input: They fed the AI data about how atoms are arranged (distances between them).
The Goal: Predict the "formation energy" (how stable the crystal is).
The Comparison: They pitted their new "Manta Ray + Magic Leap" AI against other famous search methods (like Particle Swarm Optimization, which mimics bird flocks, and Genetic Algorithms, which mimic evolution).

The Results

The paper claims their new method, EELM-MRFO-LF, was the winner.

Accuracy: It predicted the energy of the crystals with very high precision (almost matching the slow, expensive supercomputer simulations).
No Outliers: It didn't make wild, crazy mistakes on new, unseen data.
Speed: It found the best solution faster and more reliably than the other methods because the "Magic Leap" (Lévy Flight) kept the search from getting stuck in local traps.

Summary

The paper presents a new way to train an AI to predict crystal structures. By teaching the AI to search like a Manta Ray, but giving it the ability to make giant, random jumps (Lévy Flight) when it feels stuck, the researchers created a tool that is faster and more accurate at finding the "perfect recipe" for new materials than previous methods.

What the paper does NOT claim:

It does not claim this has been used to cure diseases or treat patients.
It does not claim this will immediately build new batteries or phones (though it helps design the materials for them).
It strictly focuses on predicting energy levels in binary crystal systems (Lithium and Germanium) to improve the speed of material discovery.

Technical Summary: Evolutionary Extreme Learning Machine of ab-initio Energy Landscapes for Crystal Structure Prediction using Manta Ray Optimization with Lévy Flight

Problem Statement
Predicting the crystal structure of materials based on constituent elements is a computationally intensive task, particularly when utilizing traditional Density Functional Theory (DFT) to determine the atomic configuration that minimizes total energy. While Machine Learning (ML) offers a pathway to accelerate structure search and predict unseen data with lower computational costs, standard Extreme Learning Machines (ELMs) face specific limitations. Although ELMs are faster than traditional gradient descent methods and offer improved generalization, they often require a high number of hidden neurons and can suffer from ill-conditioned problems due to the random selection of input weights and hidden biases. Furthermore, standard ELMs may converge to local minima. To address these issues, metaheuristic approaches are often employed to optimize input weights, yet existing algorithms like the Manta Ray Foraging Optimization (MRFO) can suffer from slow convergence precision and a tendency to become trapped in local optima.

Methodology
The paper proposes a novel hybrid methodology termed Evolutionary EELM-MRFO-LF, which integrates an Extreme Learning Machine (ELM) with a modified Manta Ray Foraging Optimization algorithm enhanced by Lévy Flight (LF).

Dataset and Representation: The study utilizes a dataset of approximately 14,000 Li-Ge crystal structures generated via a Genetic Algorithm for Structure Prediction (GASP) and evaluated using DFT. The input data is represented using Partial Radial Distribution Functions (RDF) to capture atomic distances, while the target output is the formation energy (both unrelaxed and relaxed) relative to pure components.
The ELM Framework: The core model is a Single-Layer Feedforward Network (SLFN). The ELM randomly initializes input weights and hidden biases. The output weights are analytically determined using the Moore–Penrose (MP) generalized inverse of the hidden layer output matrix.
Optimization Strategy (MRFO-LF): To overcome the random initialization drawbacks of standard ELMs, the MRFO algorithm is used to select the optimal input weights and hidden biases. MRFO mimics three foraging behaviors: chain foraging, cyclone foraging, and somersault foraging.
- Lévy Flight Integration: A key modification involves incorporating Lévy Flight trajectories specifically during the somersault foraging phase. This is designed to increase the diversity of the population of search agents (ELM configurations). By introducing a random walk based on a heavy-tailed distribution, the algorithm enhances global exploration and mitigates the risk of premature convergence to local optima.
Training Process: The algorithm iteratively updates the position of "manta rays" (representing weight vectors) to minimize the Root-Mean-Square-Error (RMSE) between predicted and DFT-calculated energies. The fitness function is calculated using the analytical output weights derived from the MP inverse.

Key Contributions

Novel Hybrid Algorithm: The paper introduces EELM-MRFO-LF, a specific integration of Lévy Flight into the MRFO algorithm tailored for training ELMs.
Enhanced Exploration: By applying Lévy Flight during the somersault foraging stage, the method explicitly targets the diversification of search agents, aiming to escape local optima more effectively than standard MRFO or other metaheuristics.
Application to Crystal Structures: The methodology is applied to the prediction of formation energies in binary systems (specifically Li-Ge), addressing the challenge of predicting both unrelaxed and relaxed energy states relative to ground state structures.

Results
The performance of EELM-MRFO-LF was evaluated against several state-of-the-art techniques, including ELM optimized by Genetic Algorithms (ELM-GA), Whale Optimization Algorithm (ELM-WOA), Particle Swarm Optimization (ELM-PSO), traditional MRFO (ELM-MRFO), and basic ELM. The evaluation used 5-fold cross-validation over 20 random experiments with 300 hidden nodes.

Accuracy: The proposed EELM-MRFO-LF achieved the highest performance metrics. For unrelaxed formation energy prediction, it attained an average RMSE of 9.300 meV/atom and an $R^2$ of 0.995. For relaxed formation energy, it achieved an RMSE of 9.100 meV/atom and an $R^2$ of 0.996.
Comparison: These results outperformed ELM-PSO (9.350 meV/atom for relaxed), ELM-MRFO (10.39 meV/atom for relaxed), and significantly outperformed the basic ELM (19.80 meV/atom for relaxed).
Generalization: The results indicate that combining MRFO with Lévy Flight improves not only the data fitting capabilities but also the generalization properties of the ELM, reducing outliers in the prediction of unseen data across the compositional space.

Significance and Claims
The paper claims that the EELM-MRFO-LF strategy provides a superior testing performance compared to traditional MRFO and other metaheuristic-based ELMs. The primary significance lies in the algorithm's ability to iteratively improve exploration capabilities and avoid local optima through the diversification provided by Lévy Flight, while simultaneously improving the convergence properties of the MRFO algorithm. The authors assert that this approach offers a balanced trade-off between training and testing performance, providing a robust machine learning tool for accelerating the search of energy configurations in materials science without the computational cost of repeated DFT calculations. Future work is noted to focus on applying this strategy to other engineering optimization problems in materials science and developing new hybrid meta-heuristic versions of MRFO.