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

Adrian Rubio-Solis · 2026

Read Evolutionary Extreme Learning Machine of ab-initio Energy Landscapes for Crystal Structure Prediction using Manta Ray Optimization with Lévy Flight by Adrian Rubio-Solis (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17148.txt.

Introduction

Rubio-Solis improves manta-ray foraging optimization with Lévy flights and uses it to train extreme learning machines, single-hidden-layer networks whose output weights come from a generalized inverse. The method predicts formation energies for binary crystal compounds, relaxed and unrelaxed, against their pure-component ground states. Tests compare it with other nature-inspired algorithms under matched conditions.

Catalog record

CREATORAdrian Rubio-Solis
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28715
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.17148

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Evolutionary Extreme Learning Machine of ab-initio Energy Landscapes for Crystal Structure Prediction using Manta Ray Optimization with Lévy Flight Adrian Rubio-Solis Abstract The Manta Ray Foraging Optimization algorithm (MRFO) has proven to be a powerful heuristic strategy in the optimal solution of a large number of engineering problems. In this paper, an improvement of MRFO with Lévy Flight is suggested for the training of extreme learning machines (ELMs) whose basic model is a Single Layer Feedforward Network (SLFN). The proposed methodology that we called Evolutionary EELM-MRFO-LF for short is implemented to the prediction of unrelaxed and relaxed formation energy compounds relative to ground state crystal structure of pure components in binary systems. EELM-MRFO-LF follows the learning procedure of traditional Evolutionary ELMs in which first MRFO with LF is used to select the input weights and Moore–Penrose (MP) generalized inverse is applied to analytically determine the output weights. Lévy Flight trajectory is implemented for increasing the diversity of the population of ELMs against premature convergence and the ability of avoiding getting trapped in a local optima. The performance of the suggested EELM-MRFO-LF is compared with other well-known nature-inspired algorithms under similar conditions. I Introduction Extreme Learning Machine (ELM) was originally introduced as a new class of off-line learning methodology for the training of Single-hidden-Layer Feedforward Neural Networks (SLFNs), much faster and suitable to provide improved generalisation capabilities than traditional gradient descent learning approaches [ 1 , 3 , 4 , 5 ] . ELM randomly selects the input weights and hidden biases of a SLFN and analytically determines its output weights through simple generalized inverse operation of the hidden layer output matrix. Moreover, ELM overcomes some drawbacks that face traditional gradient descent approaches such as stopping criteria, learning rate, learning epochs, and local minima [ 2 , 3 ] . However, it is also found that ELM usually tends to require more hidden neurons than traditional gradient-based learning algorithms as well as sometimes resulting in ill-condition problems due to the random selection of input weights and hidden biases [ 6 , 7 ] . To address this problem, numerous studies in the literature have proposed metaheuristic-based approaches as a mechanism to select input weights and biases and Moore–Penrose (MP) generalized inverse to analytically determine the output weights [ 3 , 4 , 5 , 8 , 11 , 9 , 10 ] . For example in [ 6 ] , an evolutionary ELM with linear hidden neurons and using particle swarm optimization (PSO) demonstrated that ELM approach is able to achieve good generalization performance with much more compact networks. PSO was applied to optimize the input weights and hidden biases of SLFNs to address prediction tasks, with boundary conditions primarily incorporated into the PSO framework while enhancing the performance of the ELM. In [ 12 ] , authors proposed an enhanced ELM method to the modelling of monthly river flow from hydro-climatic data using reptile search optimization (RSA). The proposed method was also compared to a number of popular metaheuristics, including salp swarm algorithm (SSA) and equilibrium optimizer (EO). Results demonstrated an improved model accuracy of traditional ELM in the prediction of river flow using popular accuracy metrics such as Root-Mean-Square-Error (RMSE), Mean-Absolute-Error (MAE), Nash-Sutcliffe Efficiency (NSE) and Determination Coeffcient ( R^{2} ). Metaheuristic-based selection of input weights and biases improves the generalization performance of ELM and mitigates convergence to local minima. MRFO has demonstrated its capability to solve a wide range of real world problems [ 13 ] . It is also known, MRFO suffers from slow convergence precision and it is easily trapped in a local optimal [ 27 ] . In [ 27 ] , To overcome this, authors implemented Tent chaotic mapping, a bidirectional search and Lévy Flight. While Tent chaotic mapping distributes more uniformly the initial solutions, bidirectional search expands the search area. From results presented [ 27 ] , incorporating Lévy Flight clearly contributed the algorithm’s ability to escape from local optimal. In this paper, a new method combining Extreme Learning Machine and Manta Ray Foraging Optimization with Lévy Flight (MRFO-LF) and called EELM-MRF-LF for short is suggested for the prediction of formation energies in crystal structures relative to the pure components across the entire compositional and configurational space of binary systems. As described in [ 16 ] , predicting the crystal structure of a material given its constituent elements plays a crucial role to determine the total energy of the structure. Especially, predicting the atomic configuration that minimises the total energy usually results computational intensive using traditional techniques such as Density Functional Theory (DFT) [ 16 , 17 ] . Machine learning has become an attractive computational tool not only to accelerate the structure search by recognising energy configurations with a much lower cost than DFT, but also to accurately predict unseen data [ 16 ] . The proposed EELM-MRFO-LF extends the work presented in [ 16 ] , by introducing a new machine learning strategy that provides an improved prediction accuracy and reduces the number of outliers. EELM-MRFO-LF is applied to iteratively determine the optimal parameters of SLFNs. A data set of 14,000 Li-Ge structures was employed. As indicated in [ 13 ] and [ 14 ] , MRFO usually suffers from slow convergence precision and is easily trapped in a local optima. To address this limitation, LF was incorporated into the MRFO algorithm to enhance its ability to escape local optima. [ 14 ] . In this work, EELM-MRFO-LF adopts LF to increase the diversity of search agents during somersault foraging ensuring that the global search achieves a higher exploration performance and avoids getting trapped in a local optima. To evaluate the performance of EELM-MRFO-LF with respect to other popular metaheuristics, techniques such as Whale Optimization Algorithm (WOA) [ 24 ] , continuous Genetic Algorithm (GA) [ 23 ] , Particle swarm optimization (PSO) [ 2 ] and traditional ELM [ 2 ] were also implemented. The rest of this paper is organised as follows. In section II, background material is revisited and the proposed EELM-MRFO-LF is also described. In section III, experiments and discussions are presented, while in section IV conclusions are drawn. II METHODS II-A Dataset In this paper, the data set generated in [ 16 ] is used to test the proposed evolutionary ELM using Manta Ray Foraging Optimization with Lévy Flight (EELM-MRFO-LF). Such data consists of approximately 14,000 Li-Ge structures generated by a Genetic Algorithm for structure prediction called GASP for short [ 17 ] . As described in [ 16 ] , GASP generates an initial population of structures and employs DFT to evaluate their associated energy. Relevant energy for structure predictions as a function of composition in the formation energy relative to the crystal of pure components, i.e.: E_{f}=E_{tot}-X_{Li}E_{Li}-X_{Ge}E_{Ge}, (1) in which, E_{tot} is the total energy per atom of the Li-Ge crystal structure, X_{Li} and X_{Ge} are the molar fractions of Li and Ge in the structure. E_{Li} and E_{Ge} are the energy per atom of pure Li and Ge. For comparison reasons, this paper the term ’unrelaxed’ formation energy ( E_{f}^{u} ) is employed to denote the energy of unrelaxed structures, while the term relaxed formation energy ( E_{f}^{r} ) refers to the energy

Read the complete text