Predicting how atoms pack into a stable crystal means searching a rugged energy landscape with thousands of local valleys. Modern pipelines run many independent search workers, often hundreds of short "basin-hopping" chains. Basin hopping means repeatedly perturbing a structure, relaxing it to the nearest valley, and keeping the result if it is good enough. To make each step affordable, the pipelines score structures with a machine-learned surrogate for quantum-mechanical energy. Two kinds of waste follow. Workers keep rediscovering the same valleys because nothing records where others have been. They also keep falling into false valleys that exist only because the surrogate is wrong, because when one worker's expensive density functional theory (DFT) check exposes a surrogate error, no other worker hears about it.
We propose to connect the workers through the environment rather than through messages or a central coordinator, the way ants coordinate through pheromone trails. Each worker deposits persistent, slowly decaying marks into a shared field over a low-dimensional structural descriptor space. One kind of mark says "this basin is already known." A second says "DFT disagreed with the surrogate here." Every worker adds the local field value to the energy it uses when deciding whether to accept a move. The claim is concrete. On chemistries where the surrogate is known to mislead, the distrust channel should recover ground states that surrogate-ranked search misses. Across the board, the occupancy channel should cut redundant relaxations, so the number of workers can grow without duplicated effort.
Core Claim
The unit of coordination is a small DFT audit: a worker relaxes its current best structure with DFT and compares the result with the surrogate's prediction. That audit is the most expensive and most informative event in the search. Today its result is used once and thrown away. We claim that turning each audit into a lasting, spatially spread-out penalty lets a fleet of independent workers avoid surrogate blind spots. The fleet should do this at DFT budgets of a few dozen calls, a regime where retraining the surrogate on so few points is unreliable. Stigmergy (coordination through persistent traces left in a shared medium) supplies the mechanism. Evolutionary and heuristic structure search supplies the setting and the benchmarks.
Concretely, worker $i$ judges a candidate structure $x$ by
$$ \tilde E(x) = E_{\mathrm{sur}}(x) + \lambda \sum_{j} a_j\, e^{-(t-t_j)/\tau_a}\, K\big(\phi(x),\phi_j\big) + \mu \sum_{k} d_k\, e^{-(t-t_k)/\tau_d}\, K\big(\phi(x),\phi_k\big). $$
The terms are as follows:
- $E_{\mathrm{sur}}$ is the surrogate energy per atom.
- $\phi(x)$ is a structural descriptor of dimension about 8.
- $K$ is a Gaussian kernel over descriptor distance.
- Occupancy deposits $j$ are placed at descriptor $\phi_j$ and time $t_j$, with strength $a_j$ equal to 1 whenever a relaxation lands in a basin already in the shared record.
- Distrust deposits $k$ carry strength $d_k = \max(0, E_{\mathrm{DFT}} - E_{\mathrm{sur}})$ from an audit.
- $\tau_a$ and $\tau_d$ are separate decay times.
- $\lambda$ and $\mu$ weight the two channels.
The field is stored as a list of deposits rather than a grid, so its cost grows with the number of deposits, not exponentially with descriptor dimension. This answers the substrate-dimensionality problem that the research file flags for stigmergy in abstract spaces. Workers only read and append to this list, asynchronously, and never exchange state.
How This Sits Against Prior Work
Several papers motivate the gap. Surrogate exploitation is documented directly. Multi-Objective Quality-Diversity for Crystal Structure Prediction (MOME-X) fills a property-indexed archive with Pareto fronts scored by CHGNet and MEGNet, and reports that search drives candidates into structures that fool these surrogates. Its archive stores elites passively: a discovered artifact does not change how nearby cells are explored. Integration of generative machine learning with the heuristic crystal structure prediction code FUSE (FUSE2) seeds basin hopping with a generative graph network ranked by proxy potentials. It reports that on YWB$_4$ the proxy ranking actively hindered the search, and its basin-hopping trajectories share no memory. Crystal Structure Prediction of Binary Alloys via Deep Potential relaxes about 20 million Al-Mg structures under particle swarm search. There the only shared state is a centralized global-best vector, and the discarded structures leave no trace.
Other papers supply the method. The idea of coordinating through decaying, spreading tuples injected into a shared medium comes from Programming stigmergic coordination with the TOTA middleware. We move it from network topologies to a structural descriptor space. Deep reinforcement learning for multi-agent coordination shows virtual pheromones resolving congestion among agents competing for the same narrow passages, which is the physical analogue of many workers crowding into one basin. Validating a Termite-Inspired Construction Coordination Mechanism Using an Autonomous Robot shows a purely local rule (deposit when sensed humidity falls below a threshold) producing coherent structure without a blueprint. Our workers likewise react only to the local field value. EvoGit uses a Git repository as the asynchronous shared substrate for LLM agents, a precedent for append-only records as coordination media. Adding a penalty to a primary objective to keep search physical follows A. J. Hanson et al., which combined diffraction R-factors with interatomic energies.
One paper constrains the framing. Self-Organization and Artificial Life: A Review names a gap between the top-down fitness evaluation of evolutionary algorithms and bottom-up self-organization. We do not claim to close it. Energy remains a global objective. What becomes decentralized is the memory of where the search has been and where the surrogate is untrustworthy. We also do not claim novelty for depositing repulsive kernels in a low-dimensional space; that idea has long relatives in bias-deposition methods for molecular simulation. The new elements are the second channel fed by DFT disagreement, the independent decay of each channel, and asynchronous use across a CSP worker fleet.
Research Program
We will build the field as a lightweight service inside a FUSE-style modular basin-hopping engine, following The Flexible Unit Structure Engine (FUSE). FUSE assembles cation-anion sub-modules and uses ionic-radius checks to discard overlapping candidates. CHGNet is the default surrogate, VASP handles DFT audits, and each worker is a basin-hopping chain.
Two descriptor choices will be compared. The first is hand-built: a sorted coordination-number histogram per species together with volume per atom. The second is learned: CHGNet's pooled structure embedding reduced by principal component analysis to 8 dimensions. Deciding whether two relaxed structures are "the same basin" uses a symmetry-aware structure matcher. Audits are triggered by a fixed rule: each worker audits its current best structure every $N$ accepted moves, subject to a global DFT cap.
The analysis has three parts. First, we will measure how often relaxations land in already-known basins, with and without the occupancy channel, as the number of workers grows from 8 to 256. Second, we will map where distrust deposits accumulate and check whether they concentrate in the regions MOME-X and FUSE2 identify as surrogate failures. Third, we will sweep the decay times $\tau_a$ and $\tau_d$. The research file lists decay calibration as the central open problem for stigmergic search. We expect distrust to need slow decay, because surrogate error is a fixed property of the surrogate. We expect occupancy to need fast decay, because mild re-exploration can pay off. The sweep tests this directly.
Baselines
The naive reference point is the fleet of independent random-start basin-hopping workers with CHGNet ranking and the same DFT cap, which is the FUSE2 random-start configuration. The strongest heuristic competitor on the same benchmark is FUSE2's ML-seeded search. On its 11-compound suite, FUSE2 reports an average speedup of about 2.2x over random starts, 8.3x on WCl$_2$, and an energy 338 meV/atom lower than random starts on Si$_6$Al$_6$B$_3$Fe$_3$NaO$_{30}$F. We will run our field both on random starts and on top of FUSE2's seeds, with identical wall-clock time, surrogate calls, DFT cap, and seeds. Beating it means a lower median time-to-ground-state across the suite, and a DFT ground-state recovery on YWB$_4$ where FUSE2's ranking failed.
For search efficiency on a fixed composition, the reference is oblivious local search from Crystal Structure Prediction via Oblivious Local Search. It found the Y$_2$Ti$_2$O$_7$ optimum after 720 visited structures, while basin hopping had not found it within 3500. We will count structures visited in the same unit and use the same GULP force field. Beating it means fewer than 720 visited structures in aggregate across workers, or matching 720 with a wall-clock gain from parallelism.
For diversity, MOME-X is the competitor. It reports a median of 3.0 to 7.0 matches to Materials Project reference structures across C, Si, SiC, SiO$_2$, and TiO$_2$. We will run its exact surrogates and evaluation budget, with our field added as a penalty on its mutation acceptance. Success means more DFT-confirmed reference matches and a lower fraction of archive elites that DFT rejects.
Three ablations isolate the key ingredients. The first uses the occupancy channel only. The second uses the distrust channel only. The third is a "hard tabu" version: the same deposits, but applied as fixed exclusion radii with no kernel spreading and no decay, maintained by a central coordinator that updates all workers synchronously. That last ablation tests whether the stigmergic properties, rather than shared memory as such, matter.
Why not just…?
Why not just fine-tune the surrogate on the DFT audits? This is the most serious threat. Online fine-tuning uses the same information as the distrust channel and could correct the landscape everywhere, not just near audit points. Our bet is that at 20–60 DFT points, fine-tuning a universal potential is unstable: it can distort regions far from the new data, and each update forces every worker to reload the model. The field changes the landscape only locally and immediately. We will run online fine-tuning with the identical audit stream and DFT cap, and also a combined version. If fine-tuning matches or beats the distrust channel at every budget, we will drop that channel's claim. The contribution then shrinks to the occupancy channel and to the field as a cheap, stable stopgap between fine-tuning rounds.
Why not just penalize surrogate uncertainty with an ensemble? An ensemble penalty needs no DFT at all. However, ensemble members trained on the same data tend to share blind spots. The YWB$_4$ failure reported in FUSE2 was a confident misranking by a generic proxy, not flagged uncertainty. We will train a five-member CHGNet ensemble and compare its disagreement with realized DFT error at every audit. If the ensemble penalty works as well as the distrust channel on YWB$_4$ and the MOME-X systems, the distrust channel is unnecessary.
Why not just keep a central tabu archive of visited basins? A centralized record with exclusion radii may capture most of the benefit, and the particle swarm in Crystal Structure Prediction of Binary Alloys via Deep Potential already shows central state sharing at scale. The hard-tabu ablation settles this. If it matches the full field at small fleets but falls behind at 256 asynchronous workers, the stigmergic design earns its place through scaling. If it matches everywhere, we will report the simpler method as the finding.
Decisive Experiments
The benchmarks are the FUSE2 11-compound suite, with YWB$_4$, WCl$_2$, and the borosilicate as critical cases; the Y-Sr-Ti-O phase field, including the 88-atom Y$_2$Ti$_2$O$_7$ pyrochlore cell; and the MOME-X five-system suite scored against Materials Project. Metrics are:
- DFT energy of the best structure relative to the known ground state,
- surrogate calls and DFT calls to reach it,
- fraction of relaxations landing in already-known basins,
- fraction of DFT audits that reject a surrogate-selected structure,
- MOQD-SCORE and reference-match counts for the MOME-X suite.
Each configuration will be run over at least six seeds per compound, the FUSE2 protocol scale, with medians and bootstrap intervals reported.
Three results would decide the proposal. On YWB$_4$, the distrust channel must reach the DFT ground state more often than FUSE2 seeding and more often than the ensemble penalty, at a matched DFT budget. As the fleet grows to 256 workers, the full field must keep the rediscovery fraction roughly flat while independent workers show it rising. That scaling curve must also separate the full field from the hard-tabu ablation. Across all benchmarks, the distrust channel must match or beat online fine-tuning at DFT budgets under 60 calls. If any of these fails, the corresponding part of the claim is withdrawn.
References
- Multi-Objective Quality-Diversity for Crystal Structure Prediction
- Integration of generative machine learning with the heuristic crystal structure prediction code FUSE
- Crystal Structure Prediction of Binary Alloys via Deep Potential
- Programming stigmergic coordination with the TOTA middleware
- Deep reinforcement learning for multi-agent coordination
- Validating a Termite-Inspired Construction Coordination Mechanism Using an Autonomous Robot
- EvoGit: Decentralized Code Evolution via Git-Based Multi-Agent Collaboration
- Optimization of genetic algorithm techniques for powder structure solution
- Self-Organization and Artificial Life: A Review
- The Flexible Unit Structure Engine (FUSE) for probe structure-based composition prediction
- Crystal Structure Prediction via Oblivious Local Search