Basin-Hopping Searches

• Basin-hopping searches are global optimization techniques defined by iterative random perturbations followed by local descent to navigate rugged energy landscapes.
• They are widely used in computational chemistry and materials science with extensions such as population-based, parallel, and adaptive schemes that enhance performance.
• Advanced adaptations incorporate chemically-informed moves and surrogate-assisted relaxations to reduce computational cost and improve search reliability in complex systems.

Basin-hopping searches are robust global optimization methods designed to efficiently explore rugged energy landscapes characterized by multiple local minima separated by barriers. Originating in physical chemistry for cluster structure search, basin-hopping (BH) algorithms have become widely used in numerical optimization and computational materials science due to their simplicity, reliability, and extensibility. The core principle involves stochastic perturbations of current solutions followed by local minimization on the underlying objective function, with subsequent acceptance or rejection governed by a Metropolis-type criterion. Modern adaptations include population-based variants, parallel implementations, surrogate-assisted relaxations, and chemically meaningful curvilinear trial moves.

1. Fundamental Algorithmic Structure

The canonical BH procedure interleaves random global moves with deterministic local descent. At each iteration, a random perturbation is applied to the current configuration $x$, yielding $y = x + \sigma$, where $\sigma$ is typically drawn uniformly or from a Gaussian distribution. The point $y$ is then locally minimized—using solvers such as L-BFGS-B or conjugate gradient—to reach the nearest local minimum $z$. This sequence is interpreted as a trajectory over the basins of attraction of the objective function $f(x)$, effectively transforming the landscape into a plateau-based topology where barriers are mapped to discrete jumps. The update is accepted according to the Metropolis criterion: $$P_\text{accept} = \begin{cases} 1, & \Delta f \le 0\ \exp\left(-\frac{\Delta f}{T}\right), & \Delta f > 0 \end{cases}$$ with $\Delta f = f(z) - f(x)$ and $T$ a temperature parameter controlling uphill moves (Baioletti et al., 2024). This mechanism facilitates escape from deep local minima and enables exploration across widely separated basins.

2. Extensions: Population-Based, Parallel, and Adaptive BH

Population-based variants such as BHPOP maintain a set of locally minimized solutions, enabling competitive replacement and promoting diversity. Elitist steady-state schemes replace the worst member when a better basin is found, and generational strategies allow local competition and periodic restarts to mitigate stagnation (Baioletti et al., 2024). Parallel BH implementations exploit simultaneous evaluation of multiple perturbed candidates, utilizing multi-core or distributed architecture for wall-clock speedup. For atomic clusters, parallelization achieves near-linear acceleration up to 8 cores for LJ potentials, and is extensible to ab initio relaxations via distributed computing frameworks (DFT job arrays, MPI, or workflow systems) (Carmona et al., 29 Oct 2025). Adaptive schemes dynamically tune perturbation scales based on observed acceptance rates (e.g., multiplicative or exponential updates), maintaining efficient exploration/exploitation balance and avoiding random-walk dynamics.

Variant	Mechanism	Benchmark Outcome
BHPOP	Steady-state population, restarts	Robust on multimodal landscapes
Parallel BH	Simultaneous local minimizations	~7–8× speedup up to 8 cores
Adaptive stepsize	Acceptance-rate-controlled scaling	Stabilizes global search

3. Chemically-Informed and Surrogate-Assisted Moves

Recent developments utilize trial moves in chemically meaningful coordinates. Curvilinear basis sets—delocalized internal coordinates (DICs), complete DICs (CDICs)—are constructed via SVD of the metric on primitive internal coordinates (bonds, angles, dihedrals). Moves in DIC space preserve molecular connectivity, enable constrained sampling (e.g., freezing bond stretches), and allow explicit rigid-body translations/rotations for subsystem-specific motion. For adsorbate/surface systems, CDIC moves facilitate uniform lateral site sampling and discovery of conformational diversity inaccessible in Cartesian moves (Krautgasser et al., 2016). This approach reduces the rate of dissociation and expedites local optimization by producing more chemically reasonable geometries.

Surrogate-assisted BH leverages machine-learned local relaxations—chiefly sparse Gaussian process regression based on SOAP descriptors. By substituting computationally expensive DFT minimizations with surrogate relaxations for most steps, only following up with brief DFT refinements periodically, the number of ab initio calls is reduced by up to an order of magnitude while retaining exploration effectiveness (Rønne et al., 2022). Surrogates can be pre-trained (transfer learning) or continuously updated via mini-batch k-means sparsification, facilitating concurrent multi-stoichiometry searches with a single model.

4. Algorithmic Innovations: Skipping, Grids, and Random Walk over Basins

The BH-S algorithm replaces conventional Gaussian perturbations with linear skipping proposals from rare-event sampling. Instead of a single displacement, the algorithm “skips” along a chosen direction, iteratively extending until entering a sublevel set (a basin of lower energy than the current state). This non-local mechanism connects distant basins directly, outperforming conventional BH on test surfaces with widely separated minima—although it is less efficient on landscapes with extremely narrow global attractors (Goodridge et al., 2021). Hybrid schedules alternating BH and BH-S steps may be effective in such cases.

Adaptive grids discretize the configurational space but remain bias-free due to flexible relaxation of unoccupied sites under fictitious pair potentials. Moves consist of occupied–vacant site swaps, with post-move grid optimization maintaining uniform site coverage. The grid reduces the effective search dimensionality; single-atom swaps combined with local grid relaxations have demonstrated significant reductions in accepted steps needed to reach global minima for challenging Lennard-Jones clusters (e.g., GM₃₈ found in ~3500 steps versus standard BH requiring tens of thousands) (Paleico et al., 2020).

For discrete systems such as spin glasses, basin-hopping is implemented via a random walk over basins of attraction, partitioned into energy bins for finer sampling. Minima identification is handled by steepest-descent classification, and ridge-descent procedures are used to estimate energy barriers between discovered basins. The algorithm efficiently constructs disconnectivity graphs up to N ≈ 100 spins and achieves near-exact recovery of barrier heights and overlap distributions consistent with theoretical predictions in the thermodynamic limit (Zhou, 2011).

5. Performance Benchmarks, Comparative Analyses, and Applicability

Comprehensive benchmarks on synthetic landscapes (BBOB suite) and real-world energy minimization problems (Lennard-Jones, Morse clusters) establish BH as competitive with, and often superior to, established metaheuristics—especially differential evolution (DE) and particle swarm optimization (PSO) (Baioletti et al., 2024). On short evaluation budgets, BH and BHPOP perform robustly, whereas covariance matrix adaptation evolution strategy (CMA-ES) surpasses BH only under tighter convergence criteria or at large numbers of evaluations. Empirical results in atomistic simulation (LJ clusters up to N=100, surface–adsorbate systems, molecular conformer searches) confirm that enhanced BH approaches discover global minima efficiently, with reduced dissociation rates and lower local optimization cost per accepted move.

Surrogate-accelerated BH demonstrates a ~2–10× reduction in DFT calls for molecular and cluster optimization, robust generalization to larger systems via transfer learning, and applicability to concurrent composition space searches (Rønne et al., 2022). Adaptive grid BH yields unbiased global minima recovery for heterogeneous interfaces and large supported clusters (Paleico et al., 2020). DIC/CDIC moves enable efficient sampling of conformational, tautomeric, and adsorption-site diversity (Krautgasser et al., 2016). BH-S skipping proposals dramatically increase reliability and efficiency on “distant basin” benchmarks (Goodridge et al., 2021).

6. Parameterization, Limitations, and Future Perspectives

Critical parameters include temperature $T$, perturbation scale $s$ (typically 5–20% of domain width), population size $N$ (in BHPOP), grid multiplier $N_\text{mult}$, and chemical coordinate mode selection (in DIC/CDIC BH). Overly small perturbations induce stagnation, while excessively large scales cause random-walk dynamics. Adaptive scaling and restart-triggered diversity maintain robust searching. For parallel BH, synchronization bottlenecks limit scaling beyond seven to eight concurrent minimizations (Carmona et al., 29 Oct 2025). In surrogate BH, model update cost ($\mathcal{O}(NM^2)$), prediction cost ($\mathcal{O}(M)$), and sparsification efficiency require careful tuning.

Limitations include overhead in coordinate SVD and grid relaxation in curvilinear and grid-based methods, incomplete coverage for exceedingly high-dimensional or combinatorial spaces, and inefficiency of skipping proposals for landscapes with infinitesimal global basin measure. Possible extensions involve fully asynchronous BH acceptance schemes, PID-like controllers for step-size adaptation, hybridization with covariance matrix updates (as in CMA-ES), integration with GPU-accelerated local minimizations, application to noisy or constrained objective spaces, and broader benchmarking on mixed-variable and multi-objective functions.

In summary, basin-hopping searches constitute a versatile and robust toolkit for global optimization in both continuous and discrete settings, with proven efficacy in atomistic structure determination, numerical benchmarking, and chemically complex landscapes. Their extensibility—from parallel and population-based mechanisms to surrogate and chemically-aware trial moves—ensures continued relevance for large-scale, heterogeneous, and multi-speciality optimization tasks.