Generative models now design metal-organic frameworks (MOFs, porous crystals built from metal clusters joined by organic linker molecules) by placing whole building blocks rather than individual atoms. About two thirds of what they produce fails basic structural checks, and the failure rate rises as the number of blocks per unit cell grows. Physicists who study granular matter have a cheap tool for exactly this kind of question. They count how many constraints a network of contacts imposes against how many ways its pieces can move, and that count tells them whether a packing is floppy, just barely rigid (they call this "marginally stable"), or over-braced. This note proposes to carry that counting up to the level of building blocks. Each linker or metal cluster becomes a rigid body. Each block-to-block joint becomes a contact whose type, "frictionless" or "frictional," depends on whether the joint can transmit torque. The resulting block-level rigidity analysis would be cheap, defined in the generator's own coarse coordinates, and smooth enough to steer generation.
The central bet is that the coarse-grained level is where these mechanical failures are both decided and visible. If floppy block-level networks predict which generated frameworks fall apart on relaxation, after controlling for size, then a generator can be steered away from them before any atoms are placed. If they do not predict failure, the idea is wrong in a clean, checkable way.
Core Claim
Granular physics supplies the counting rule and its known failure mode. Frictionless spheres jam when the contacts exactly balance the degrees of freedom. Studies of compression and shear jamming show this isostatic point carries universal scalings (Criticality and marginal stability of the shear jamming transition of frictionless soft spheres; Marginal Stability Constrains Force and Pair Distributions at Random Close Packing). Adding tangential, frictional forces at contacts makes the network over-constrained, or hyperstatic, and naive counting breaks (Unifying Suspension and Granular flows near Jamming; Entropy of Jammed Matter). We reuse this distinction rather than suffer from it.
In a coarse-grained MOF, a linker that binds a metal cluster through a single coordinating atom acts like a frictionless pin. It fixes three translational constraints and leaves rotation free. A chelating or bridging carboxylate, or a rigidly fused joint, acts like a frictional contact that also resists rotation. With $N_b$ rigid blocks in a periodic cell and joints $j$ each removing $c_j \in \{3,5,6\}$ degrees of freedom, the Maxwell-style count of floppy modes is
$$N_{\text{floppy}} \ge 6N_b - 3 - \sum_j c_j ,$$
where the $-3$ removes rigid translations of the whole cell. As in the full analysis, equality fails when the geometry is special, so the count is only a lower bound. Our analysis therefore does not stop at counting. It assembles a block-level rigidity matrix, of size roughly $(\sum_j c_j) \times 6N_b$ instead of atoms-by-atoms, and diagonalizes it.
The atomistic study we build on, Singularly isostatic and geometrically unstable rigidity of metal-organic frameworks, built UFF4MOF spring networks for the 5,682 CoRE 2019 MOFs. It sorted them into generically rigid, singularly isostatic, and geometrically unstable classes, and showed that porous frameworks cluster near the isostatic threshold. It also showed that symmetric geometries create "accidental" zero modes: UiO-66 has 238 at the Γ-point. Weak long-range constraints added through a cutoff τ lift these modes into soft, flat bands. Our claim is that most of this classification can be recovered from the block-level framework once joints carry the right torque-transmission type and weak inter-block springs stand in for τ. We further claim that the block-level spectrum predicts which coarse-grained generator outputs fail.
One contradiction has to be faced directly. UiO-66 is a real, robust material that is "geometrically unstable" in the harmonic model. So zero modes alone cannot be the definition of invalidity, and a generator should not be pushed toward maximal rigidity. The target is the subtler class of networks that remain floppy even after the weak auxiliary constraints are added. We expect generator failures, not real materials, to concentrate there.
Research Program
The first step is the block-level rigidity model. We take the building-block decomposition used by MOFDiff, which diffuses metal nodes and linkers as rigid units in a periodic cell with fixed lattice parameters and then assembles atoms by overlap matching. Each block-to-block connection point receives a joint type, assigned from the local coordination chemistry by a simple rule set: monodentate gives a pin, bidentate gives a hinge or full weld depending on geometry. We then add weak inter-block springs within a cutoff, the coarse analogue of τ. From this we compute the block-level rigidity matrix, its zero-mode count, and the low end of the block dynamical-matrix spectrum, and from those assign each structure to one of the three rigidity classes.
The second step is validation against atomistic ground truth on CoRE 2019. This answers whether a model with about a hundredfold fewer degrees of freedom reproduces the atomistic classes, and in particular whether it catches accidental modes from symmetric nodes. The joint-type ablations matter here. If an all-pin, purely frictionless model matches the atomistic classes as well as the chemistry-typed model, the torque-transmission idea adds nothing.
The third step is diagnosis of the generator. We apply the block-level analysis to MOFDiff samples before atom assembly. We then test whether residual floppiness, meaning zero modes that survive the auxiliary springs, predicts failure of UFF relaxation or MOFChecker, beyond what block count and lattice size already explain.
The fourth step, attempted only if the third succeeds, is guidance. Rank is discontinuous, so we use a smooth surrogate: the sum of $\exp(-\lambda_k/\epsilon)$ over eigenvalues $\lambda_k$ of the block dynamical matrix with auxiliary springs, where $\epsilon$ sets the softness. This approximates the number of residual soft modes. Its gradient with respect to block poses is added during the late denoising steps. The guidance target is the CoRE distribution of rigidity classes, not zero floppiness.
Isolated drug molecules from MOTIFLOW-style motif decompositions (3D Molecule Generation from Rigid Motifs via SE(3) Flows) are deliberately left out of scope. Acyclic motif trees are trivially floppy, with one mode per rotatable bond, so counting says nothing new there.
Baselines
The naive reference is unguided MOFDiff on BW-DB, retrained or reused at the published configuration. MOFDiff reports a validity rate of about 30% on BW-DB (MOFDiff), so this is the number to beat. We will also report validity as a function of block count, since the paper describes a steep drop with size.
The strongest competitor for compute-efficient validity is MOFDiff with post-hoc filtering using the full atomistic rigidity analysis of the CoRE study, combined with UFF relaxation and MOFChecker. It sees the true atomistic structure and needs no new theory. We run it under equal wall-clock budgets with identical checkers. Guidance beats it only if valid, non-duplicate frameworks per GPU-hour, including all filtering costs, rise by a clear margin with matched novelty and uniqueness. Matching validity at lower cost would count as a partial win only.
The third baseline is a graph neural network trained to predict the atomistic rigidity class from the block graph, used both as a filter and as a classifier-guidance signal. It isolates whether the physics is needed at all, or whether a learned predictor on the same coarse inputs does the job.
The ablations isolate the key ingredient. One assigns every joint as a pin, the pure frictionless limit. One assigns every joint as a weld, the fully frictional limit. One removes the auxiliary springs. One replaces the spectral surrogate with the bare Maxwell count. There is no reported result on Lacuna for block-level rigidity classification of MOFs, so the atomistic CoRE labels are the only reference for that experiment.
Why not just…?
Why not just filter generated samples with the atomistic rigidity matrix or UFF relaxation? This is the most serious threat. If block-level floppiness merely duplicates what a post-hoc filter finds, then filtering is simpler and at least as accurate. Our case rests on two things filtering cannot do: catching failures before atom assembly, and changing what the sampler proposes. Only if guided sampling raises yield per unit compute does this proposal earn its keep. The equal-budget comparison in the experiments settles this directly, and we will report a loss honestly if filtering wins.
Why not just guide with force-field energy? Energy guidance rewards locally relaxed geometry. Floppy mechanisms are by definition energy-neutral, so a structure can have low UFF energy and still carry many collapse modes. GEOM-Drugs Revisited also shows that classical force fields can mask geometric errors that a quantum-level check exposes. We include UFF-energy guidance as an arm to test this, and expect it to improve local geometry without changing the rigidity-class distribution.
Why not just train a classifier on rigidity labels? A learned predictor may match the physics model on structures that resemble CoRE. It has no reason to generalize to block topologies or joint chemistries it never saw, and generators produce exactly such out-of-distribution structures. We test this by holding out entire topology families from classifier training. If the classifier matches the physics model on held-out topologies too, we will say so and recommend it as the cheaper tool.
Proposed Experiments
Experiment one runs the block-level analysis on all decomposable CoRE 2019 structures. It reports agreement with the atomistic three-class labels, recall on known accidental-mode cases such as UiO-66, and the speedup over atomistic diagonalization, broken out across the joint-type ablations.
Experiment two is the decisive check. On several thousand MOFDiff samples we fit a logistic model of relaxation and MOFChecker failure against block count, cell volume, and block-level residual soft-mode count. If the soft-mode coefficient is not significant once size is controlled, or adds negligible predictive power on held-out samples, the guidance stage is abandoned. In that case the result stands as a negative finding: MOFDiff's size-dependent failures are not mechanical in origin.
Experiment three, conditional on experiment two, compares guided and unguided MOFDiff and the baselines on BW-DB. It measures validity, uniqueness, novelty, validity versus block count, and how closely the rigidity-class distribution matches CoRE. A secondary check asks whether guidance preserves the CO₂-capacity optimization reported for MOFDiff, which needed about 14 instead of about 58 expensive gas-adsorption simulations to find frameworks above 2 mol/kg (MOFDiff). The point is to confirm that stabilizing the frameworks does not destroy their function.
References
- Criticality and marginal stability of the shear jamming transition of frictionless soft spheres
- Marginal Stability Constrains Force and Pair Distributions at Random Close Packing
- Unifying Suspension and Granular flows near Jamming
- Entropy of Jammed Matter
- Singularly isostatic and geometrically unstable rigidity of metal-organic frameworks
- MOFDiff: Coarse-grained Diffusion for Metal-Organic Framework Design
- 3D Molecule Generation from Rigid Motifs via $\mathrm{SE}(3)$ Flows
- GEOM-Drugs Revisited: Toward More Chemically Accurate Benchmarks for 3D Molecule Generation