Aether 1.0 — a simulation-native foundation model

Abstract

We introduce Aether 1.0, a foundation model trained primarily on simulation trajectories, instrument data and design history rather than on scraped natural-language text. Aether is intended as a unified substrate across fluid, structural, electromagnetic, molecular, quantum, biological and digital-circuit dynamics. We train three model sizes on a shared corpus, evaluate against held-out experimental endpoints across every discipline, and demonstrate substantive cross-domain transfer between disciplines. We describe the architecture, training procedure, evaluation methodology, capability-gating safety mechanism, and the failure modes we ship the model with. All evaluation cases and reproduction scripts are open.

1. Introduction

Chat-trained foundation models read about the physical world. They are excellent at describing it and weak at simulating it. The reasons for this are well-understood: the training objective rewards predicting the next text token, the corpus is overwhelmingly natural-language prose, and the inductive biases of the architecture have been tuned for sequence modelling rather than dynamics. The result is a model that can write a competent paragraph about turbulence-modelling choices but cannot integrate the Navier-Stokes equations forward in time on a new geometry.

This is not a criticism of chat-trained models. It is an observation about what a different training objective could accomplish. We hypothesised that a model trained on the trajectories of real physical systems — coupled solver runs across fluid, structural, electromagnetic, molecular, quantum and digital dynamics — would develop a latent state corresponding to governing dynamics rather than to language. Whether such a model would transfer across disciplines, and whether it would do so well enough to be useful in practice, was an open empirical question.

Aether 1.0 is our answer. The model is trained primarily on simulation, secondarily on instrument data and design history, and tertiarily on engineering text — in that order of weight. It runs forward in time across physical systems, calls numerical solvers as tools, and ships with capability-gated safety controls suitable for autonomous-lab deployment. We release three model sizes: a 7B dense, a 40B dense, and a 280B sparse mixture-of-experts with 40B active parameters per token.

In this paper we describe the corpus (Section 2), the architecture (Section 3), training procedure (Section 4), evaluation methodology and headline results (Sections 5–8), safety architecture (Section 9), known failure modes (Section 10), and a reproducibility appendix (Section 11). Numbers and reproduction scripts are open for every result claimed below.

2. The corpus

Aether's pretraining corpus is built from four sources, weighted by physical-fidelity utility rather than by raw token count. The first source is solver runs — petabytes of computational-fluid-dynamics, finite-element analysis, molecular dynamics, density-functional theory and digital-RTL simulation trajectories, computed on internal infrastructure across thousands of cases. The second source is instrument traces — plate-reader curves, mass-spectrometry runs, high-content imaging stacks, manufacturing-line sensor streams, and post-tape-out silicon telemetry. The third source is design history — geometry files, layout databases, netlists and parametric design records for products that were actually shipped, where licensing permits us to use them. The fourth source is engineering text from authoritative standards bodies and textbooks; we treat this as grounding rather than as a substitute for the data above.

We upweight failure archives. Incident reports, recall databases, root-cause analyses and postmortem documents are over-sampled during training because they encode the boundary between what designs work and what designs do not — a signal that is otherwise under-represented in a corpus dominated by successful runs. Empirically, this is the single most impactful curation decision we made; ablating the upweighting reduces downstream rare-event-prediction quality by a substantial margin.

We do not train on customer data. The corpus is internal, licensed, or open. Customer prompts, code and experiment data submitted to the platform are never used to train base models — a property we enforce contractually and audit through immutable trace records of training-data lineage.

3. Architecture

Aether uses a transformer architecture with three modifications relative to a vanilla autoregressive language model. First, the tokenisation accommodates non-text modalities natively — mesh nodes, instrument channels, and circuit elements are tokenised by structural role rather than by lexical surface form. Second, the attention pattern is grid-aware: for inputs with explicit spatial structure (CFD grids, FEA meshes, layout floorplans), attention is biased toward spatial neighbours rather than purely positional ones. Third, the model has a dedicated tool-use head, separate from the next-state-prediction head, that emits structured tool calls into the runtime.

The three published sizes — 7B dense, 40B dense, 280B sparse-MoE — share the same vocabulary, the same training corpus and the same evaluation suite. The sparse-MoE configuration routes per token through 32 experts of 8B parameters each, activating 5 experts per token for an effective active-parameter count of 40B. Active-parameter inference cost is therefore comparable between the 40B dense and the 280B sparse-MoE; the larger model wins by carrying more diverse specialisations across its experts.

Context length is 1M tokens across all configurations. The window is genuinely usable: we benchmark performance at the full 1M length and demonstrate retrieval and reasoning at the high end of the window. The mechanism that makes 1M feasible is a combination of ring attention, sequence-parallel attention and a careful KV-cache layout that allows streaming long-horizon rollouts without recomputing the full state on each new token.

4. Training

Training uses a multi-task objective: next-state prediction (the largest term), a conservation-law regularisation penalty (described in our long-rollout-stability paper), a tool-use cross-entropy term, and a refusal-corpus regression loss that enforces capability-gating behaviour. The conservation-law term is soft — the model can violate conservation when the data requires it, but the gradient signal is large enough to materially change the learned dynamics.

We trained the 7B, 40B and 280B sparse-MoE configurations on the same corpus at fixed-equivalent compute budgets to enable fair scaling-law analysis. The scaling-law fits and exponents are documented in detail in our scaling-laws paper; the headline is that simulation pretraining returns a steeper loss-vs-compute exponent (≈0.087) than text-trained baselines at comparable architectures.

We employed several training-stability interventions. We mixed sequence lengths within batches rather than padding to a fixed length, used learning-rate warmup over the first 1% of total tokens, and applied a cosine decay to the minimum LR rather than to zero. We saw no loss spikes during the published runs after the first 200B tokens, in line with what well-instrumented frontier training runs report.

5. Evaluation methodology

Our evaluation philosophy is straightforward: measure against measured reality, not against other models. Each discipline has a panel of held-out experimental cases — wind-tunnel data, plate-reader assays, taped-out silicon, measured material properties — and we report against the experimental endpoints rather than against a held-out training split. For benchmarks where measured data does not exist (rare but real for some classes of multiphysics workload), we report against high-fidelity reference simulations with documented physics provenance.

We treat every published number as reproducible. The case, the mesh, the solver settings, the extraction script and the random seed are stored alongside the result. Customers on the Research tier can re-run any benchmark by pulling the reproduction kit. We treat unreproducible numbers as bugs and have a public regression log.

We separate two evaluation axes: single-step accuracy and long-rollout stability. They are not the same. A model with lower single-step cross-entropy can have worse long-horizon stability if its errors compound in a particular way. We track both axes per release and publish both.

6. Headline results — engineering

On the standard impinging-jet benchmark, the 40B dense model holds within 3.6% mean Nusselt error across radial stations, against measured experimental data. The 280B sparse-MoE configuration improves on this to 3.4%. The leading commercial RANS baseline at matched mesh and time-step budget comes in at 7.8% on the same benchmark.

On the standard transonic wing-body benchmark at cruise conditions, the model produces lift polar deviations under 0.6% across the full angle-of-attack range, drag polar deviations under 1.0%, and shock-location prediction within 1.2% of measured chord position. This is competitive with the best results in the published literature on fully-resolved structured meshes, obtained on a coarser unstructured mesh.

On notched-bar fatigue (standard ASTM cases), the model produces life predictions with Pearson correlation 0.92 against measured cycles-to-failure across the published series. The remaining error is concentrated in the very-low-cycle regime, where plasticity-driven life prediction remains hard for all approaches we are aware of.

7. Headline results — drug discovery and lab

On the public free-energy-perturbation benchmark across eight protein targets, Aether Bio 12B (a specialised variant pretrained additionally on biology and chemistry trajectories) achieves an aggregate RMSE of 1.06 kcal/mol against measured relative binding affinities. The per-target breakdown is documented in our FEP benchmark paper; the model is strongest on kinase targets and weakest on large allosteric binders.

On hERG-channel inhibition classification, ROC-AUC reaches 0.91 on an external curated set; with customer-assay calibration the AUC rises to 0.94. On aqueous solubility (logS) regression, RMSE is 0.45 log units on an external set — competitive with specialised endpoint models.

On the public antibody-developability flag panel, aggregate ROC-AUC against measured downstream failure outcomes is 0.83. Aggregation propensity is the strongest single endpoint at 0.89; viscosity at therapeutic-formulation concentrations is the weakest at 0.71, due to insufficient use of solvent and pH context in the model's current implementation.

8. Headline results — semiconductors

On the public open-cell-library timing benchmark, Aether for semiconductors' physical-implementation flow (driven by Aether) closes worst-negative-slack within 0.4% of the commercial-EDA baseline at matched density. On the public 7nm predictive PDK benchmark suite, the flow achieves 1.06× performance at iso-area against a leading commercial baseline. EM/IR signoff results agree with the leading commercial tool within 2.0% on current density across the published test set.

Where the commercial baselines still win is on designs with heavy clock-domain crossings, where they hold a small but consistent edge. We are working on a domain-aware planning head; current results are documented honestly rather than presented as parity where parity does not yet exist.

9. Headline results — materials, quantum and cosmology

On materials discovery, Aether Materials 12B (a specialised variant pretrained additionally on DFT trajectories and lattice-dynamics data) achieves formation-energy MAE of 36 meV/atom across the public Materials Project test split — competitive with the best dedicated graph-network baselines on the same split. Phonon-dispersion prediction reproduces Γ-X-M-R band shapes within visual fidelity on 87% of held-out compositions; the remaining 13% are concentrated in heavy-element systems with strong spin-orbit coupling where the training corpus is thin.

On quantum-circuit simulation, the model drives a variational quantum eigensolver to chemical accuracy (within 1 mHa of full-CI ground state) on small molecules under realistic NISQ noise models. We simulate up to 64 qubits with the model's combination of tensor-network compression and noise-aware ansatz selection. The same model proposes QAOA circuits for portfolio-optimisation toy problems and discovers known-good operator orderings without being told to.

On cosmology, Aether Earth 18B (pretrained additionally on N-body trajectories and weak-lensing maps) reproduces ΛCDM halo-mass functions within 4% across the 10¹² – 10¹⁵ solar-mass range against reference 1024³-particle N-body simulations. Weak-lensing emulation hits the survey-cosmology required accuracy on the Aemulus benchmark suite. The civilisational-scale Earth-system variant is in pilot with three partner institutions.

10. Headline results — autonomous labs

On the Cell Tracking Challenge public dataset, the Apex world model achieves R²=0.79 on cell-count prediction and Dice score 0.84 on instance segmentation across the released image series. Performance is calibrated on real microscopy across multiple imaging modalities; reproduction kits ship with the case definitions, the model checkpoint hash and the seeds.

Active-learning campaigns running on real instruments demonstrate Bayesian-optimal designs that move the posterior more efficiently than vendor schedulers at matched plate budget. On a calibrated benchmark across five assays, Aether active-learning campaigns reach a target performance threshold in approximately half the plate volume of the vendor baseline. Empirical 90% credible intervals achieve 92% measured coverage with conformal calibration enabled.

ALCOA+ and 21 CFR Part 11 audit posture is verified by the autonomous-lab discipline; electronic batch records are auto-written as campaigns run and are signed by agent ID, instrument ID and model version. Chain-of-custody is enforced by the runtime, not by post-hoc reconciliation. Biosecurity refusals on controlled-pathogen scenarios are verified against the same versioned red-team corpus that gates the model release.

11. Headline results — software engineering

On the SWE-bench Verified benchmark, the Aether software-engineering configuration resolves 53.6% of issues with whole-repo planned edits — a number we audit against the original benchmark scripts and the human-verified test split. On the broader SWE-bench Full split, resolution is 38.7%. We hold the headline number against the verified set rather than the easier full set, and we publish failure-class analysis alongside the win.

Cross-language migration on six target languages (TypeScript, Python, Go, Rust, Java, Swift) preserves test-suite semantic equivalence on 94% of migrated modules against the source-language tests, on the public Crosslang-700 benchmark we built and released. The remaining 6% are concentrated in pattern-matching and reflection-heavy modules where the target language has no clean translation; we flag rather than silently translate these.

CVE-aware review at PR time identifies known-pattern vulnerabilities (CWE top 25) with precision 0.91 and recall 0.83 on a curated 12k-PR audit set drawn from open-source repositories. The same model proposes the patch in 64% of detection cases; the patch passes the original test suite without modification in 71% of proposed cases.

12. Cross-domain transfer

The most surprising empirical result from the Aether training programme is the strength of cross-domain transfer. Checkpoints with high exposure to CFD during pretraining measurably improve downstream fine-tuning performance on molecular dynamics, on protein-ligand docking, and — most unexpectedly — on cell-painting image analysis. The effect is statistically robust across multiple training runs and across multiple downstream task formulations.

Our working hypothesis is that the shared latent across these tasks is governing dynamics — conservation laws, continuity, advection, diffusion — that apply across spatial scales. A model that has learned to predict the evolution of a turbulent flow has learned, at some level of abstraction, that mass and energy are conserved and that stuff moves smoothly. That abstraction is useful for predicting the evolution of cell morphology under a drug stimulus, even though the surface forms of the two tasks are radically different.

The effect is not universal. Pretraining on RTL digital simulation does not measurably help docking or ADMET — we suspect because digital state evolution is discrete in a way that fluid and biological state are not. Pretraining on quantum chemistry helps materials and crystal-structure prediction (as you would expect from the underlying continuity) but the transfer to fluid dynamics is small. The pattern is consistent: continuous-dynamics → continuous-dynamics transfers; discrete → continuous is weak.

13. Safety architecture

Capability gating in Aether lives in the model's tool-use policy rather than in a pre-prompt or a post-filter. Refusal scenarios are encoded in the action distribution itself, so an attempt to elicit a refused action by prompt engineering, by reformulating into an adjacent domain, or by chaining tool calls does not bypass the refusal. The mechanism is described in detail in our biosecurity-boundaries paper; the headline is that the refusal is a property of the trained model rather than a property of careful prompt construction by the customer.

We verify the refusal corpus against a versioned red-team set before every model release. The corpus is hash-versioned alongside the model — every release runs against the corpus version pinned at release time. Customers in regulated environments can pin a corpus version independently from the model version and extend it with their own scenarios; the verification result flows back to us as an aggregate signal without leaking the corpus contents.

External red-team passes run quarterly. The findings drive corpus updates and, occasionally, architectural changes. We publish summary statistics from each red-team pass and acknowledge contributing researchers publicly on /security/hall-of-fame.

11. What 1.0 is not yet good at

We publish failure modes alongside successes. Five categories of weakness in Aether 1.0:

• Highly novel chemistries far outside the training distribution — particularly transition-metal complexes with unusual coordination spheres and main-group compounds with low literature coverage. The model is over-confident on these and we are working on uncertainty calibration that captures the out-of-distribution character.
• Rare-event reliability in small-sample regimes. We under-predict tail risk in some accelerated-fatigue cases and in extreme-load crash simulations where the training corpus has few examples. The fix is more data on these tails, plus a conditional sampling approach that we will document in 1.1.
• Anything our curation pipeline under-represents. Materials physics for sub-3nm semiconductor nodes is a known gap because the training corpus does not yet include enough validated data on those nodes. Laminar-to-turbulent transition in hypersonic boundary layers is another known weakness.
• Time-dependent CYP inhibition in ADMET. The training distribution under-represents time-dependent inhibition relative to its true clinical importance. The model under-predicts time-dependent inhibitors at the high end.
• Antibody viscosity at therapeutic-formulation concentrations. The model under-uses solvent and pH context; we are training a small auxiliary head on solvent + pH and the preview results are encouraging but the published numbers reflect the current state.

12. Limitations and future work

Aether 1.0 is a foundation model, not a complete solution to scientific computing. Customers deploying it will encounter cases that are out-of-distribution for the corpus, workflows we have not yet built agents for, and integration requirements that need custom forward-deployed engineering. We document these honestly in pilot kickoff meetings; we treat the customer's hardest cases as input to our training-data prioritisation rather than as cases we tell the customer to work around.

Future work is documented on /changelog and prefigured in the failure modes above. Aether 1.5 is planned for late 2026; its focus areas are the known weakness list and broader coverage of semiconductor PDKs. Beyond that, we expect specialised configurations to grow — additional smaller variants for edge deployment, additional larger variants where economics support frontier capability.

13. Reproducing this work

All benchmark cases referenced in this paper are open. The simulation cases, mesh files, solver settings and extraction scripts are part of the platform; customer access for reproduction is free under the Research tier. We provide a reproduction kit per benchmark that includes the case definition, the model checkpoint hash, the random seeds, and a one-command runner. Cases run in roughly the wall-clock time documented for each one; the longest case in the public set takes approximately 12 hours on an 8-GPU node.

If reproduction fails, write to research@apexworldlabs.com. We treat reproduction failures as bugs and we maintain a public regression log of issues filed, root-caused and fixed.

Acknowledgements

Aether 1.0 is the work of the full Apex technical team. We thank the customers who allowed us to use their workloads as pilot cases and who pushed back on early checkpoints; we thank the external red-team researchers whose findings drove substantial corpus updates; and we thank our compute partners, whose flexibility on scheduling and capacity made the 280B training run feasible inside our planned timeline.