[Paper Review] A Brief Review on Results and Computational Algorithms for Minimizing the Lennard-Jones Potential

Jiapu Zhang|arXiv (Cornell University)|Dec 30, 2010

Graph theory and applications33 references3 citations

TL;DR

This paper presents a hybrid optimization approach combining discrete gradient and simulated annealing methods to minimize the Lennard-Jones potential energy for atomic clusters. By approximating the nonconvex potential with a piecewise linear function to reduce local minima, the method achieves globally optimal or near-optimal solutions for N up to 308 atoms, with results matching or improving upon known benchmarks for many configurations.

ABSTRACT

The Lennard-Jones (LJ) Potential Energy Problem is to construct the most stable form of $N$ atoms of a molecule with the minimal LJ potential energy. This problem has a simple mathematical form $f(x) = 4\sum_{i=1}^N \sum_{j=1,j<i></i>

Motivation & Objective

• To address the challenge of minimizing the nonconvex Lennard-Jones potential energy function, which has exponentially many local minima with increasing N.
• To develop a computational method that effectively navigates the complex energy landscape of atomic clusters to find the global or near-global minimum energy configuration.
• To improve upon existing results for the Lennard-Jones potential by reducing the number of local minima through function approximation.
• To provide a robust and scalable algorithm applicable to large atomic clusters, particularly in protein folding and materials science contexts.
• To validate the method through numerical experiments and compare results with known optimal values for N ≤ 308.

Proposed method

• The Lennard-Jones potential function is approximated using a piecewise linear function g(τ) = max(g₁(τ), min(g₂(τ), g₃(τ))) to reduce the number of local minima.
• The approximation divides τ into three segments: (0, r₀], [r₀, r₁], and [r₁, 16], with r₀ = ∛2 and r₁ = 1/∛(2/7), ensuring accurate representation of the original φ(τ) = τ⁻⁶ - τ⁻³.
• A hybrid algorithm is employed: first, the discrete gradient method is applied with a build-up technique to generate an initial solution.
• Then, the simulated annealing method is used with the approximated objective function to escape local minima and refine the solution.
• The method alternates between discrete gradient and simulated annealing phases, starting from the improved initial solution to converge to a global minimum.
• The algorithm is tested on clusters of size N from 19 to 308, with results compared against known benchmarks.

Experimental results

Research questions

• RQ1 Can a piecewise linear approximation of the Lennard-Jones potential effectively reduce the number of local minima in the energy landscape?
• RQ2 How does the hybrid discrete gradient and simulated annealing method compare to existing global optimization techniques in solving the Lennard-Jones problem for large N?
• RQ3 To what extent can the proposed method improve upon previously known optimal energy values for N ≥ 17?
• RQ4 Does the use of a build-up technique combined with nonsmooth optimization enhance convergence to the global minimum?
• RQ5 Can the method reliably achieve near-optimal solutions for atomic clusters up to N = 308?

Key findings

• The method achieved the best-known energy value for N = 19, 20, 21, 24, 34, 49, 56, 65, 67, 84, 93, 148, 170, 172, 268, 288, 293, 298, 300, 301, 304, and 308, matching or improving upon prior results.
• For N = 22, the method obtained -86.573675, slightly below the known best value of -86.809782, indicating a potential improvement.
• For N = 27, the result of -112.825517 was obtained, slightly below the known best value of -112.873584, suggesting a marginal improvement.
• For N = 30, the computed value of -128.096960 was below the known best value of -128.286571, indicating a notable improvement.
• For N = 65, the result of -334.014007 was below the known best value of -334.971532, showing significant improvement.
• For N = 308, the method achieved -1999.983235, closely matching the known best value of -1999.983300, demonstrating high accuracy for large clusters.