HeteroSynth: Modeling Structured Heterogeneity

• HeteroSynth is a conceptual framework that explicitly models heterogeneity across diverse applications such as materials synthesis, inverse scattering, polymer copying, and spiking neural networks.
• It integrates generative models like conditional diffusion for synthesis planning with low-rank latent representations to balance precision and diversity in multi-modal design spaces.
• The approach couples statistical learning with domain-specific constraints, achieving phase discrimination, improved fidelity, and efficiency across physical, chemical, and computational systems.

HeteroSynth is a term used in several technically distinct research settings to denote systems that explicitly model heterogeneity rather than averaging it away. In current arXiv usage, it refers to an envisioned AI framework for heterogeneous materials synthesis planning, concretely realized for zeolites by the diffusion model DiffSyn; a synthetic benchmark for inverse subsurface scattering in heterogeneous media; a heterogeneous templated-copying platform in nonequilibrium polymer physics; and, in spiking neural-network research, the concept named HetSyn, where synapse-specific time constants implement heterogeneous temporal integration (Pan et al., 21 Sep 2025, Tiwari et al., 4 Sep 2025, Guntoro et al., 2024, Deng et al., 1 Aug 2025). Across these settings, HeteroSynth denotes an operational commitment to structured variability in composition, interaction energies, optical fields, or temporal constants.

1. Scope of the term

Current usage is not standardized to a single artifact. The label appears in multiple literatures with different ontological status—framework, dataset, physical platform, or naming variant of a neural model.

Context	Object referred to as HeteroSynth	Core task
Materials informatics	Envisioned AI system; DiffSyn is a concrete realization	Generate synthesis recipes conditioned on target structure and templating agent
Inverse rendering	Synthetic dataset	Infer heterogeneous scattering parameters from sparse multi-view images
Nonequilibrium polymer physics	Heterogeneous templated-copying platform	Study copying speed, fidelity, and efficiency under heterogeneity and detachment
Spiking neural networks	Same concept as HetSyn	Learn synapse-specific timescales for temporal computation

A common misconception is that HeteroSynth denotes a single software package or benchmark. The available literature instead uses the name for multiple domain-specific constructs. Adjacent methods are also explicitly positioned as components of a broader HeteroSynth-style workflow: Hetero2d for 2D-substrate screening, chemical heteroencoders for chemically meaningful latent spaces, and DeepRHP for random heteropolymer design (Boland et al., 2021, Boland et al., 2021, Li et al., 10 Jun 2026).

2. Heterogeneous materials synthesis planning

In the materials-synthesis literature, HeteroSynth is formulated as an AI system that plans synthesis routes for heterogeneous materials by modeling the probabilistic, multi-modal mapping from desired structures to feasible recipes. DiffSyn is the concrete realization of this idea for zeolites: a conditional generative diffusion model trained on ZeoSyn, a corpus of 23,961 synthesis routes extracted from 3,096 papers spanning 1966–2021, covering 921 unique OSDAs, 233 zeolite topologies, and 1,022 unique materials (Pan et al., 21 Sep 2025).

The generation task is explicitly conditional:

$$p(x_{\mathrm{comp}}, x_{\mathrm{cond}} \mid c_{\mathrm{zeo}}, c_{\mathrm{OSDA}}),$$

where the model predicts probable gel compositions and reaction conditions given a target framework and a chosen organic structure-directing agent. DiffSyn jointly models 12 continuous synthesis variables, including heteroatom and ion ratios such as Si/Al, Al/P, Si/Ge, Si/B, Na$^+$/T, K$^+$/T, OH$^-$/T, F$^-$/T, H$_2$O/T, template content SDA/T, and crystallization temperature and time. Because the empirical parameter distributions are non-Gaussian, these variables are quantile-transformed to a uniform distribution with sklearn.preprocessing.QuantileTransformer, stabilizing learning over diverse and multi-modal regimes.

Conditioning is implemented through separate encoders for the zeolite and the OSDA. Zeolites are represented by invariant geometric descriptors such as pore volume, ring sizes, largest included sphere, and $FD_{\mathrm{Si}}$; OSDAs are represented by physicochemical descriptors including molecular volume, shape axes, SASA, charge, NPR ratios, rotatable bonds, PMI, and sphericity. These embeddings are fused and injected into a conditional DDPM via classifier-free guidance. The diffusion core follows the standard forward process

$$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$

with reverse denoising learned by a conditional U-Net and guided at sampling time by

$$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$

using $p_{\mathrm{uncond}}=0.1$ and $^+$0. The model uses input dimension 12, downsampling widths 128, 64, and 32 with mirrored upsampling, $^+$1 diffusion steps, exponential moving average with $^+$2, batch size 32, learning rate $^+$3, and 1M epochs; training on a single NVIDIA RTX A5000 took approximately 51 hours. An EGNN encoder with $^+$4, $^+$5, and $^+$6 did not outperform invariant descriptors, with Wasserstein 0.60 versus 0.53, so invariant features were retained by default.

DiffSyn is motivated by the one-to-many and multi-modal nature of structure–synthesis relationships. For the AEL framework, principal component analysis of literature recipes shows two dominant modes; regression models and GANs collapse to one, while normalizing flows and VAEs recover both modes but with many false positives. DiffSyn was reported to capture the true multi-modal distribution with high-quality, diverse samples, achieving the lowest Wasserstein distance, outperforming the VAE by over 25%, achieving higher precision than other generative models, and the lowest MAE for 10 of 12 parameters despite not being trained on MAE. It also reproduces joint dependencies such as the inverse correlation between temperature and crystallization time consistent with the Arrhenius relation

$^+$7

and learned domain heuristics such as H$^+$8O/T versus $^+$9, with Spearman correlations $^+$0 for H$^+$1O/T versus $^+$2 and $^+$3 for temperature versus $^+$4.

A central HeteroSynth capability is phase discrimination. DiffSyn delineated the phase boundary between FAU and LTA in the OSDA-free regime, and predicted an analogous boundary for ERI versus KFI by generating dense samples in synthesis space associated with each phase. A second capability is objective-aware route selection: for CHA with TMAda, DiffSyn-generated routes were ranked by precursor cost and crystallization time, and Pareto-front analysis showed that low-cost and fast-crystallizing solutions occupy different parts of synthesis space; some generated routes improved on the 20 least expensive literature recipes by simultaneously lowering cost and time.

The UFI case study supplies experimental validation. DiffSyn generated 1,000 synthesis routes for UFI conditioned on Kryptofix 222, a framework–OSDA system absent from training. The generated distributions suggested high Na$^+$5/Si, low K$^+$6/Si, H$^+$7O/Si around 15, moderate F$^+$8/Si, and a temperature distribution with a major mode at $^+$9–$^-$0 and a minor mode at $^-$1. Four syntheses were then executed at $^-$2 for 168 h with seeds under dynamic conditions, yielding UFI with measured Si/Al$^-$3 values of 13.0–19.0; 19.0 is reported as the highest yet recorded for UFI and is associated with improved thermal stability in catalytic applications. Powder XRD matched simulated UFI patterns, and SEM showed “house-of-cards” morphology. DFT rationalization used the binding-energy definition

$^-$4

with $^-$5B97X-D/def2-TZVP in ORCA v5.0.4 and SMD water solvation. Na$^-$6 was found to bind more strongly to UFI CBUs than K$^-$7, with $^-$8 kJ/mol and $^-$9 kJ/mol for Na$^-$0 on wbc and rth, versus $^-$1 kJ/mol and $^-$2 kJ/mol for K$^-$3, rationalizing the model’s recommendation of high Na$^-$4/Si and low K$^-$5/Si.

The broader HeteroSynth interpretation is explicit: DiffSyn is presented as the generative engine for heterogeneous materials synthesis planning, capable of conditioning on structure plus templating agent, modeling one-to-many and multi-modal recipe distributions, discriminating competing phases, and using physics to rationalize recommendations. Its main stated limitations are equally central to that interpretation: OSDA conditioning remains mandatory, discrete synthesis factors such as precursor identity and seeds are not yet modeled, literature bias produces spiky distributions, and sequential denoising is slower than one-step decoders.

3. Adjacent computational backbones for HeteroSynth-style design

Several methods in adjacent chemical and materials domains are described as directly supporting a HeteroSynth-style workflow, even when they are not themselves named HeteroSynth. They extend the same design logic to substrate selection, latent-space molecular representation, and random heteropolymer generation (Boland et al., 2021, Bjerrum et al., 2018, Li et al., 10 Jun 2026).

Hetero2d is an open-source Python workflow for high-throughput computational synthesis and design of 2D–substrate heterostructures. It automates interface construction under lattice-mismatch constraints, VASP input generation, job submission and monitoring, and storage of structural, energetic, and electronic outputs in MongoDB. The principal workflow, get_heterostructures_stabilityWF, comprises five FireWorks: relaxation of the 2D material, relaxation of the bulk reference, relaxation of the substrate bulk, slab construction and slab relaxation, and interface generation plus relaxation of up to 2–4 stacking configurations enumerated through Wyckoff-site matching. Screening uses the Zur and McGill coincidence-supercell search with strain tolerance below 5% along $^-$6 and $^-$7, maximum supercell area below $^-$8, and initial $^-$9-separation $_2$0. Across 50 cubic elemental substrates and four target 2D materials—$_2$1-MoS$_2$2, $_2$3-NbO$_2$4, $_2$5-NbO$_2$6, and hexagonal-ZnTe—the study reduced 400 possible combinations to 49 workflows, 123 stacking configurations, and finally 78 stable interfaces across 29 workflows. Cu, Hf, Mn, Nd, Ni, Pd, Re, Rh, Sc, Ta, Ti, V, W, Y, and Zr were reported to sufficiently stabilize formation energies, with binding energies of approximately 0.1–0.6 eV/atom and $_2$7. This defines a computational backbone for HeteroSynth in substrate-assisted synthesis.

Chemical heteroencoders provide a complementary latent-space backbone. In this literature, a heteroencoder translates between different representations of the same molecular graph—for example canonical SMILES to enumerated SMILES—so that the bottleneck cannot simply memorize one serialization. The key empirical result is that decoder enumeration dominates latent-space shaping: the Can2Enum configuration achieved the best balance between latent similarity and chemical similarity on GDB-8, with fingerprint $_2$8 and sequence $_2$9, while Can2Can remained dominated by SMILES syntax with fingerprint $FD_{\mathrm{Si}}$0 and sequence $FD_{\mathrm{Si}}$1. On five QSAR datasets, heteroencoder bottleneck vectors outperformed both autoencoder vectors and ECFP4, with average $FD_{\mathrm{Si}}$2 up to 0.80 and normalized RMSE 0.75, versus ECFP4 at $FD_{\mathrm{Si}}$3 and normalized RMSE 1.00. The same work also makes the trade-off explicit: decoder enumeration increases diversity but sharply raises wrong-molecule rates, such as 50.3% for GDB-8 Can2Enum and 65.6% for ChEMBL23 Can2Enum. Within a HeteroSynth interpretation, these models provide chemically meaningful latent coordinates for downstream generation and filtering, but they also expose the precision–diversity tension of representation-level heterogeneity.

DeepRHP extends the HeteroSynth logic to random heteropolymer design. It augments a classical VAE with a parallel feature-based decoder, forcing a shared latent variable $FD_{\mathrm{Si}}$4 to reconstruct both sequence-level information $FD_{\mathrm{Si}}$5 and chemically meaningful features $FD_{\mathrm{Si}}$6, specifically average hydrophilic–lipophilic balance over sliding windows. The hybrid ELBO is

$FD_{\mathrm{Si}}$7

The training corpus consists of random heteropolymers built from MMA, EHMA, OEGMA, and SPMA, plus membrane and globular protein sequences mapped to these monomer-equivalents. In the Aquaporin Z case, latent overlap recovered the experimentally optimal composition identified by Panganiban et al., namely RHP 4 with MMA 50, OEGMA 25, EHMA 20, SPMA 5, and also highlighted RHP 5, MMA 40, OEGMA 25, EHMA 30, SPMA 5, as a plausible alternative. This suggests that HeteroSynth-style systems can use shared latent structure to align synthetic polymer ensembles with biological target spaces without requiring exact sequence determinism.

4. HeteroSynth as an inverse-scattering benchmark

In inverse rendering, HeteroSynth denotes a synthetic benchmark for studying inverse subsurface scattering in heterogeneous media when no accepted real-world distribution exists for spatially varying optical parameters. The dataset is built on the premise that Perlin and Fractal Perlin noise provide a controllable proxy for the multi-scale heterogeneities observed in fog, smoke, clouds, tissues, and translucent solids (Tiwari et al., 4 Sep 2025).

Each sample couples photorealistic rendered images with ground-truth 3D fields for the extinction coefficient $FD_{\mathrm{Si}}$8 and volumetric albedo $FD_{\mathrm{Si}}$9. Geometry is provided by 103 VOLMAP meshes, split into 90 train and 13 test shapes, embedded in a $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$0 voxel cube of side length 50 cm. A binary occupancy mask restricts the medium to the interior:

$$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$1

Heterogeneity is generated with 3D Fractal Perlin noise,

$$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$2

using $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$3, $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$4, and $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$5. The paper specifies a modulus operation applied to the noise before scaling $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$6 by a global optical-density factor $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$7, with $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$8 for point lighting and $$q(x_t \mid x_{t-1}) = \mathcal{N}\!\left(x_t; \sqrt{1-\beta_t}\,x_{t-1}, \beta_t \mathbf{I}\right),$$9 for environment lighting. Albedo is constrained to $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$0. The phase function is Henyey–Greenstein with $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$1, so anisotropy does not vary.

Rendering uses Mitsuba 3 v3.0.1. Each object is observed from six views at $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$2 spacing around the up-axis, at $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$3 resolution and 4096 spp. Lighting is either outdoor HDRI environment illumination from Poly Haven or point lights in one-light-at-a-time configurations, with one point light co-located with the camera. Dataset scale is approximately 1.086M images, with roughly 1.08M train and 6.63k test images, plus ground-truth $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$4, $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$5, 3D occupancy masks, 2D foreground masks, and meshes.

The benchmark is used to evaluate TensoIS, a feed-forward inverse-scattering model that regresses low-rank tensor components rather than full 3D volumes. From six encoded image features $$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$6, it forms a latent

$$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$7

then reconstructs each volume through rank-$$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$8 vector–matrix outer products,

$$\tilde{s}_\theta(x_t,t,c) = (1+w)s_\theta(x_t,t,c) - w s_\theta(x_t,t,\phi),$$9

With $p_{\mathrm{uncond}}=0.1$0, this gives approximately 47.6% compression for $p_{\mathrm{uncond}}=0.1$1 volumes. Separate decoder branches predict components for $p_{\mathrm{uncond}}=0.1$2 and $p_{\mathrm{uncond}}=0.1$3. Environment lighting is estimated through learned $p_{\mathrm{uncond}}=0.1$4 spherical-harmonic coefficients, and the global scale $p_{\mathrm{uncond}}=0.1$5 through a separate regressor.

Training uses an occupancy-masked $p_{\mathrm{uncond}}=0.1$6 loss on volumes, an $p_{\mathrm{uncond}}=0.1$7 loss on scale, a light-estimation loss, and a feature-regularization term encouraging lighting-invariant latent features for the same medium under different illumination. With $p_{\mathrm{uncond}}=0.1$8, Adam at learning rate $p_{\mathrm{uncond}}=0.1$9, batch size 24, and 50 epochs, the representative environment-lighting configuration with 2D decoders and 10 components achieved MAE $^+$00 for $^+$01, MAE $^+$02 for $^+$03, MSE $^+$04 for scale $^+$05, MSE $^+$06 for SH coefficients, image-space $^+$07, and image MSE $^+$08. Multi-light training substantially improved accuracy: under point lighting, $^+$09 MAE dropped from $^+$10 to $^+$11, and $^+$12 MAE from $^+$13 to $^+$14.

HeteroSynth’s significance here is methodological rather than semantic: it supplies a controlled proxy distribution in a domain where the target distribution is unknown. Its limitations are correspondingly explicit. Fractal Perlin noise is not a physically validated distribution of real-world scattering parameters; the benchmark fixes $^+$15, omits surface reflection, and acknowledges difficulty in recovering the sharpest high-frequency structures. This suggests that HeteroSynth is best understood as a testbed for inverse-scattering architectures rather than a definitive model of natural heterogeneous media.

5. HeteroSynth in templated polymer copying

In nonequilibrium polymer physics, HeteroSynth denotes a heterogeneous templated-copying platform that combines position-dependent template–copy interactions with continuous product detachment from behind the growing tip. The problem is motivated by two facts: natural copying systems eventually detach the product from the template, and template-binding free energies of both matched and mismatched monomers are heterogeneous (Guntoro et al., 2024).

The coarse-grained state consists of a fixed template sequence $^+$16 and an attached copy $^+$17, with locality assumed at the tip. Addition and removal propensities depend only on the local tip state, and detachment is engineered so that at most two copy monomers remain bound to the template at any time. This bounded-contact condition is central: it suppresses the rugged, long-range free-energy landscapes that cause near-equilibrium subdiffusion in non-detaching templated self-assembly.

Each coarse-grained move is resolved into reversible microsteps obeying local detailed balance: monomer binding, polymerization coupled to breaking a generic bond, and tail detachment. Sequence dependence enters through $^+$18, the specific template–copy binding free energy, together with $^+$19 and $^+$20. The net persistent free-energy per incorporated monomer is sequence independent,

$^+$21

so sequence-specific binding advantages are transient and are offset by tail unbinding at the next step. Using an extension of Qureshi’s spanning-tree coarse-graining method, the effective propensities are

$^+$22

with explicit sequence-dependent forms derived from the micro rates. Two limiting regimes are especially informative. In the backward-discrimination regime,

$^+$23

whereas in the forward-discrimination regime,

$^+$24

The main dynamical result is that explicit detachment removes the subdiffusive behavior characteristic of non-detaching heterogeneous copying near equilibrium. The front position $^+$25 has diffusive mean-square displacement, $^+$26, and the growth velocity approaches zero linearly as $^+$27, without the long zero-velocity tail observed in templated self-assembly. Equilibrium occurs at

$^+$28

Heterogeneity has asymmetric effects depending on whether it is assigned to correct or incorrect pairings. Heterogeneity in correct interactions roughens the favorable landscape, increasing revisits, slowing growth, and increasing errors; error increases up to approximately $^+$29-fold were reported. Heterogeneity in incorrect interactions roughens the unfavorable landscape, which can selectively disfavor incorrect incorporation, reduce revisit counts, and improve both speed and fidelity. Typical average error reductions of 10–20% across Bernoulli template compositions were reported in such regimes. The roughness concept is formalized through variances of local free-energy increments, and because only finitely many contacts exist simultaneously, the impact of roughness remains bounded.

The framework also distinguishes thermodynamic efficiency from information-transfer efficiency. In the low-driving regime $^+$30, the thermodynamic efficiency is

$^+$31

while the information efficiency is

$^+$32

Heterogeneity in incorrect interactions can improve $^+$33 over a wide range of low drives, but $^+$34 improves only in more restricted windows, especially in backward-discrimination regimes. This difference prevents a common overstatement: improved thermodynamic performance does not automatically imply improved information transfer.

As a design platform, HeteroSynth in this literature is prescriptive. It recommends keeping correct interactions as homogeneous as possible, introducing controlled heterogeneity into incorrect interactions, enforcing detachment from behind the tip, and operating slightly above equilibrium for throughput. This suggests a physically grounded route to speed–accuracy–efficiency trade-offs in synthetic copying systems.

6. HeteroSynth as HetSyn in spiking neural networks

In spiking-neural-network research, HeteroSynth is explicitly identified with HetSyn, a framework that assigns a time constant to each synapse and shifts temporal integration from membrane potentials to synaptic currents. The concrete neuron model is HetSynLIF, an extended LIF model with synapse-specific decay dynamics (Deng et al., 1 Aug 2025).

For each synapse $^+$35,

$^+$36

so the heterogeneous timescale is located at the synapse. The membrane is a passive sum of synaptic currents minus a spike-triggered reset current:

$^+$37

with

$^+$38

If all $^+$39 and $^+$40 collapse to a shared $^+$41, the recurrence reduces to discrete-time LIF. Adding a second spike-triggered current with its own decay yields ALIF-like threshold adaptation. In this sense, HetSynLIF subsumes vanilla LIF, threshold-adaptive neurons, and neuron-level heterogeneous LIF by parameter restriction rather than by architecture change.

Training uses surrogate-gradient BPTT. A triangular surrogate is used for pattern generation and delayed match-to-sample, and a multi-Gaussian surrogate for SHD. Synaptic time constants are learned end-to-end via the exponential parameterization of $^+$42, with decay factors constrained to $^+$43. The framework was evaluated on pattern generation, delayed match-to-sample, SHD speech recognition, Sequential MNIST, TiDigits, and Ti46-Alpha. Across these tasks, HetSynLIF was reported to improve temporal processing, robustness to noise, working memory, efficiency under limited neuron resources, and generalization across timescales.

The quantitative classification results are strong. HetSynLIF achieved 92.36% on SHD, 98.93% on S-MNIST, 98.99% on TiDigits, and 96.53% on Ti46-Alpha. In delayed match-to-sample, feedforward HetSynLIF converged fastest and reached the highest accuracy among feedforward variants; recurrent HetSynLIF retained 100% test accuracy even when 80% of synaptic time constants were masked and non-trainable, maintained 100% accuracy at a 2500 ms delay without increased iterations to reach 95% accuracy, and retained nearly 80% accuracy under 20 Hz added noise. Under deletion noise and time warp on SHD, HetSynLIF degraded less than HomNeuLIF, HomNeuALIF, and HetNeuLIF.

The learned synaptic time constants also have an interpretive role. In SHD, last-layer hidden-to-output synaptic time constants were initialized from $^+$44 ms and shifted after training toward lower values with a mild long tail; 2.0% exceeded 20 ms. The paper relates this qualitatively to broad, long-tailed synaptic-time-constant distributions measured in mouse and human cortex by Campagnola et al. and others. A plausible implication is that heterogeneous synaptic kinetics supply a task-adaptive basis of fast and slow pathways without requiring explicit architectural modules for timescale separation.

The stated limitations are technical rather than conceptual. Per-synapse time constants increase memory relative to neuron-level heterogeneity, multiple $^+$45 combinations can yield similar dynamics, and hardware deployment requires per-synapse parameter support. Even so, the framework treats heterogeneity as a parameterized computational resource rather than a nuisance factor.

7. Common methodological themes and open issues

Across these literatures, heterogeneity is not treated as residual noise. In DiffSyn, the target is a calibrated multi-modal conditional distribution over synthesis variables; in inverse scattering, HeteroSynth provides explicit spatially varying optical fields; in templated copying, heterogeneity in $^+$46 is a control knob for kinetics and fidelity; and in HetSyn, synapse-specific $^+$47 values are the mechanism of multi-timescale computation (Pan et al., 21 Sep 2025, Tiwari et al., 4 Sep 2025, Guntoro et al., 2024, Deng et al., 1 Aug 2025).

A second recurring pattern is compressed representation of high-dimensional heterogeneous structure. DiffSyn jointly denoises 12 synthesis variables in a low-dimensional continuous parameterization; TensoIS reconstructs $^+$48 volumes from low-rank vector–matrix components; heteroencoders and DeepRHP force latent spaces to encode representation-invariant or chemically interpretable structure rather than raw serialization; Hetero2d reduces substrate screening to structured interface descriptors, energetic filters, and database-backed workflow provenance (Bjerrum et al., 2018, Li et al., 10 Jun 2026, Boland et al., 2021). This suggests that HeteroSynth, as a family of ideas, is closely tied to latent or low-rank parameterizations that preserve heterogeneity while remaining computationally tractable.

A third pattern is coupling statistical learning to domain constraints. DiffSyn rationalizes generated routes with DFT binding energies and preserves learned heuristics such as Arrhenius-like temperature–time relations. Hetero2d integrates vdW-corrected DFT, Bader analysis, and DOS diagnostics. DeepRHP anchors latent coordinates to HLB features rather than relying on sequence reconstruction alone. The templated-copying platform imposes local detailed balance and explicit detachment. HetSyn constrains decay factors by exponential parameterization and interprets learned time constants against empirical neurophysiology. The common logic is not merely heterogeneity, but heterogeneity constrained by physics, chemistry, or mechanism.

The main open issues are also shared in form. Distributional proxies may be too narrow: Fractal Perlin noise is only a proxy for real heterogeneous scattering, and literature-derived synthesis datasets encode anthropogenic rounding and reporting biases. Important discrete variables are often absent: DiffSyn currently omits precursor identity and seeds as modeled variables, while RHP generation requires post hoc enforcement of composition and sequence constraints. Precision–diversity trade-offs remain acute: heteroencoders improve latent chemical relevance but increase wrong-molecule decoding, and diffusion-based synthesis planning remains slower than one-step decoders. Identifiability is another recurring challenge, whether in inverse scattering, where multiple parameter volumes can produce similar appearance, or in HetSyn, where different weight–timescale combinations can yield similar dynamics (Tiwari et al., 4 Sep 2025, Pan et al., 21 Sep 2025, Bjerrum et al., 2018, Deng et al., 1 Aug 2025, Li et al., 10 Jun 2026).

Taken together, the available literature does not define HeteroSynth as a single canonical system. It defines a research orientation: heterogeneous variables are modeled explicitly, generated conditionally, and often checked against mechanistic constraints. In one domain this yields phase-selective zeolite synthesis planning; in another, a benchmark for volumetric inverse scattering; in another, a detaching copier with tunable fidelity; and in another, an SNN with synapse-level temporal specialization. The shared scientific claim is narrower than a universal framework but stronger than a naming coincidence: carefully structured heterogeneity can be algorithmically useful, physically interpretable, and, in several cases, experimentally or quantitatively validated.