Many nanoparticle clusters are built hierarchically. Small groups of particles form first, then those groups fuse into larger ones. When the building blocks pack in a way flat space cannot accommodate (icosahedra are the classic case), the growing cluster accumulates strain. Past some size it can spontaneously stop being round and become elongated. The standard theory of this transition treats the final cluster as the minimum of an energy and ignores the order in which it was put together. Yet the order of fusions plausibly decides whether, and when, a cluster breaks symmetry. That order is a rooted binary tree: leaves are particles, internal nodes are fusion events. It is inherited by the final structure in the same way a family tree is inherited by DNA. This proposal treats "how was this cluster assembled?" as a phylogenetic inference problem. We infer a probability distribution over fusion trees from the structural fingerprints that early fusions leave in the final cluster. We then ask whether conclusions about symmetry breaking change when one carries that uncertainty instead of trusting a single best tree.
The opportunity is that phylogenetics has spent two decades learning, often painfully, that picking one best tree and treating it as truth produces confident but wrong downstream conclusions when the data's signal is weak. Its methods for avoiding that mistake are now mature: posterior samplers over tree shapes, compact representations of huge sets of equally good trees, and sharp theory for when deep history is recoverable at all. None of this has been pointed at physical assembly. Assembly is a natural target because, unlike biology, a simulator gives the true history.
Core Claim
We claim three things. First, the final structure of a hierarchically assembled frustrated cluster carries recoverable information about its fusion tree. Subclusters that fused early share inherited local states, such as crystallographic orientation domains, twin-boundary placement, and strain patterns. These play the role of shared mutations, and post-fusion relaxation erodes them the way substitutions erode sequence signal. Second, a posterior over fusion trees gives calibrated answers to mechanistic questions, for instance "at which fusion did the shape order parameter first cross its symmetry-breaking threshold?" A single reconstructed tree gives overconfident and sometimes wrong answers to the same questions. Third, we make a falsifiable quantitative prediction borrowed from phylogenetic theory. Phase transitions in Phylogeny shows that, for a simple two-state inheritance model on trees, deep ancestry is recoverable from few characters only when $b\theta^2 > 1$. Here $b$ is the branching factor and $\theta$ is the per-generation correlation retained between parent and child state. If the analogy holds, deep fusion history should become unrecoverable once relaxation pushes the measured per-fusion orientation correlation $\theta$ below about $1/\sqrt{2}$ for binary fusions. Above that point it should remain recoverable even with few structural descriptors. The threshold may shift because assembly inheritance is neither two-state nor symmetric. Measuring where it actually sits is part of the contribution.
This is different from both parent fields. Geometrically frustrated self-assembly of hyperbolic crystals from icosahedral nanoparticles derives the sphere-to-cylinder transition by minimizing a deterministic elastic-plus-electrostatic energy on an idealized smooth slice. It has no notion of assembly pathway or ensemble. Its authors note that the slice choice is a drastic simplification. Conversely, the phylogenetic samplers we reuse have only ever been applied to species, viral, and cell lineages. The research file found no paper connecting the two. The superficially "obvious" bridge would be to embed trees in hyperbolic space because both fields use hyperbolic geometry. That bridge is shared vocabulary, not a shared problem, and we do not pursue it.
Research Program
The first component is a ground-truth simulator. We will build a coarse-grained aggregation model in which icosahedral particles and subclusters diffuse and fuse irreversibly. After each fusion, a short local relaxation runs under an energy with the two ingredients of Cheng, Sun and Mao: an elastic penalty for forcing icosahedral packing into flat space, and long-range electrostatic repulsion. We will also run a second, unfrustrated control system built from isotropic multi-well pair potentials, in the spirit of The emergence of bulk structure in clusters via isotropic multi-well pair potentials, where bulk-like motifs form without frustration. Every run records the true fusion tree and the cluster's shape order parameter after each fusion. Two knobs control the difficulty of inference: relaxation strength, which sets how fast inherited signal decays, and the ratio of elastic to electrostatic energy, which sets where symmetry breaks.
The second component turns structure into characters. For each particle in the final cluster we compute discrete local descriptors: orientation-domain label, local coordination motif, and strain-sign bins. We also measure the empirical per-fusion correlation $\theta$ by comparing descriptors just before and long after each fusion in the simulator. This calibration step lets us test the threshold prediction directly rather than assume it.
The third component is posterior inference over fusion trees. Fusion trees are rooted and built bottom-up by merging pairs. That is exactly the construction used by PhyloGFN: Phylogenetic inference with generative flow networks. PhyloGFN is a generative flow network, a sampler trained to draw objects with probability proportional to a reward. It builds trees by repeatedly choosing which two subtrees to merge and uses a likelihood computed by Felsenstein's dynamic programming. We reuse that sampler. We replace the biological substitution model with an inheritance model whose per-fusion transition probabilities depend on the relaxation energy released at that fusion. Large energy release means more scrambling of inherited states. This physics-informed likelihood is the proposal's key methodological ingredient. Where many fusion trees fit equally well, we compress the sampled set with the history subpartition directed acyclic graph from Representing and extending ensembles of parsimonious evolutionary histories with a directed acyclic graph. That paper's graph expanded 800 SARS-CoV-2 trees into $10^{26}$ equally parsimonious histories by swapping compatible subhistories. We use it to report how many assembly histories are consistent with a given cluster.
The fourth component is the scientific readout. For each cluster we compute the posterior distribution of the "symmetry-breaking fusion," the earliest internal node whose subcluster's order parameter exceeds threshold. The order parameter for subclusters is computed from the particles under that node. We also estimate how much of the variance in final elongation is explained by early-fusion features of the tree, such as the size balance of the first few merges, versus total cluster size alone. The equilibrium theory implies that size alone should suffice. Path dependence would show up as a large share explained by tree features.
Baselines
The naive reference point builds one tree by agglomerative clustering. We will use average-linkage clustering of particles on descriptor Hamming distance, the analog of distance-based tree building, and read every downstream quantity off that single tree with no uncertainty. It costs seconds. Anything we build must beat it on calibration, not just on accuracy, to justify itself.
The strongest conventional competitor is a Markov chain Monte Carlo sampler over trees run on the same discretized descriptors under a generic symmetric multi-state character model, as a phylogeneticist would do with MrBayes. It is the gold standard for posterior inference in phylogenetics. It will be given the same descriptors and the same wall-clock budget as our sampler, with chain length tuned on held-out simulated clusters. On the established phylogenetic benchmarks DS1 to DS8, the file reports that PhyloGFN's sampling probabilities correlate with the true posterior at a Pearson coefficient of about 0.90 to 0.99 across high-, medium-, and low-probability trees, whereas the variational baseline VBPI-GNN fell to about 0.46 on low-probability trees (PhyloGFN). Before any assembly experiment, we will check that our reimplementation reaches comparable correlation on those datasets. That way any later failure is attributable to the assembly problem, not to the sampler. No reported result was found on Lacuna for any method inferring assembly histories of nanoparticle clusters, so on the new task the bar is set empirically by the MCMC baseline.
The ablation isolates the physics-informed inheritance model. The identical generative-flow sampler is run with the generic symmetric character model in place of the energy-dependent one. Beating the MCMC baseline but not this ablation would mean the gain comes from the sampler, not from the physics. Beating the ablation would mean that modeling relaxation-dependent erasure matters. We count a win when the full method improves posterior calibration of subassembly support, the frequency with which a subcluster supported at probability p is truly present, and coverage of the symmetry-breaking fusion, both at matched compute and across at least three relaxation regimes.
Why not just…?
Why not just run the forward simulator many times and report the statistics? For the population-level question, how often hierarchical clusters of a given size are elongated, this works and requires no inference. We will report it as a sanity reference. It cannot say how a specific observed cluster was assembled, which is what matters when the cluster comes from an experiment. Simulation-based rejection over trees is also hopeless at our scale: the number of rooted binary trees grows super-exponentially in particle count. The real threat is a learned likelihood-free estimator trained on simulations. We will include a simple one, a graph network that predicts the symmetry-breaking fusion directly from the final structure. If it matches our coverage, that is a legitimate finding and narrows our claim to interpretability and the threshold test.
Why not just bootstrap the single clustering tree? Resampling descriptors and re-clustering is the cheap way phylogeneticists attach support values. It mixes estimation noise with model error and tends to be miscalibrated when signal is weak. Our calibration curves will show directly whether bootstrap support matches true subassembly frequencies. If bootstrapped clustering is calibrated in the high-$\theta$ regime, we will say so and confine the claim to the weak-signal regime.
Why not just apply the equilibrium energy theory and skip histories entirely? If final shape is fully determined by size, as the minimum-energy picture of Cheng, Sun and Mao implies, then fusion histories are irrelevant to symmetry breaking. They would matter only as a curiosity. The variance-decomposition experiment tests exactly this. A finding that early-tree features explain almost none of the elongation beyond size would show the equilibrium theory is enough for this system. We would then report the inference machinery as a tool for recovering history, not as an explanation of shape. The observation in Single-particle mapping of nonequilibrium nanocrystal transformations, that gold nanocrystals pass through kinetically chosen non-equilibrium intermediates, makes path dependence plausible. That paper studies single-crystal etching, not aggregation, so it motivates the hypothesis without settling it.
Proposed Experiments
The validation experiment reproduces PhyloGFN-level posterior correlation on DS1 to DS8 with our sampler implementation, and checks one dataset against the Tree-KL reported for PhylaFlow on DS2 (0.505, Hybrid Flow Matching in Billera-Holmes-Vogtmann Tree Space for Generative Phylogenetic Inference) as a second reference.
The first decisive check sweeps relaxation strength in the frustrated simulator. For each setting we measure $\theta$ and the accuracy of recovering subassemblies at each depth of the true tree. The prediction from Phase transitions in Phylogeny is a sharp drop in deep-node recovery near $2\theta^2 = 1$ that does not soften much with more descriptors. A gradual decline with no threshold, or a threshold far from that value, would falsify the transfer. The extension of this result to distance-based methods in Phase Transition in Distance-Based Phylogeny Reconstruction suggests the agglomerative baseline should show the same threshold, which we will check.
The second decisive check compares point estimates with posteriors on the symmetry-breaking question. Across held-out clusters, we compare how often the true symmetry-breaking fusion falls inside the 90% posterior interval with how often the single clustering tree identifies it exactly. The proposal's central claim fails if the single tree is right about as often as the posterior's coverage implies. It succeeds if, in the weak-signal regime, point estimates are confidently wrong while posterior intervals stay calibrated. This mirrors what Representing and extending ensembles of parsimonious evolutionary histories with a directed acyclic graph documents for low-signal viral data.
The third decisive check is the path-dependence measurement, run in both the frustrated and the unfrustrated control systems. We expect early fusion order to explain a meaningful share of final elongation only in the frustrated system. Equal shares in both systems would point to a generic aggregation effect rather than frustration-driven symmetry breaking. Together these checks say whether "assembly phylogenetics" recovers real history, whether its uncertainty is honest, and whether that history matters for the shape physics the field cares about.
References
- Phase transitions in phylogeny
- Geometrically frustrated self-assembly of hyperbolic crystals from icosahedral nanoparticles
- The emergence of bulk structure in clusters via isotropic multi-well pair potentials
- PhyloGFN: Phylogenetic inference with generative flow networks
- Representing and extending ensembles of parsimonious evolutionary histories with a directed acyclic graph
- Single-particle mapping of nonequilibrium nanocrystal transformations
- Hybrid Flow Matching in Billera-Holmes-Vogtmann Tree Space for Generative Phylogenetic Inference
- Phase Transition in Distance-Based Phylogeny Reconstruction