SIMCON kunststofftechnische Software GmbH¶

Company Overview¶

Business Overview¶

SIMCON kunststofftechnische Software GmbH ("kunststofftechnische Software" = plastics-technology software) was founded in 1988 in the Aachen area as a spin-off from RWTH Aachen University, one of Europe's leading engineering universities, by Dr. Paul Filz. It is registered at the Local Court of Aachen (HRB 3996) and headquartered at Schumanstrasse 18a, 52146 Würselen, Germany, with a French subsidiary (SIMCON SARL, Lyon). Registered trademarks include SIMCON, CADMOULD, VARIMOS and 3D-F. [Evidence] ¹⁰⁹⁶

Key corporate facts:

Size: 26–50 employees per the VDWF (German tool-and-mould-makers' association) member directory. SIMCON itself publishes no headcount or revenue — a private GmbH. [Evidence] ¹⁰ [Evidence of absence — financials]
Ownership: on June 18, 2024, Fortino Capital, a European growth investor, took a majority stake to fund product innovation and international expansion. CEO Dr. Bastiaan Oud: "We're shifting gears, and we've got ambitious plans." Founder Dr. Paul Filz remains associated with the company. [Evidence] ⁹
Track record: the company claims 12,000+ injection-moulding projects delivered and 35+ years of simulation development. [Evidence — vendor self-reported]⁴
Positioning: German trade directories describe Cadmould as the best-selling German injection-moulding simulation software. [Evidence — directory-sourced, not independently audited]¹⁰

Automotive Business¶

SIMCON does not report by industry, but automotive is its most visible segment. The Fortino investment announcement names Bosch, Continental, Roche and Volkswagen among its hundreds of customers; the company website shows logos including Arburg, Bosch, Continental, Diehl, L'Oréal, Roche and TE Connectivity — i.e., automotive OEM (VW), major Tier-1s (Bosch, Continental), a connector maker central to automotive electronics (TE), and an injection moulding machine maker (Arburg). No public case study pairing a named automotive customer with a quantified AI outcome was found. [Evidence] ⁹⁴ [Evidence of absence]

Engineering Business¶

SIMCON is essentially a one-platform company: it develops and sells Cadmould and its Varimos automation layer, plus simulation services on demand and training. Distribution is notable for a company this size: Cadmould and Varimos are sold through Siemens' Advanced Partner Alliance, licensed via Altair Units (Altair is now part of Siemens — altair.com redirects to Siemens' partner pages), and through regional resellers (e.g. CDG in the UK, Produsoft in Benelux) and a CAD partnership with TopSolid. SIMCON also participates in the German publicly funded AI project "KI-Opti-Pack" (AI hub for plastics packaging). [Evidence] ¹¹¹²

Role in Automotive Moulded Parts Workflow¶

Workflow Stage	Relevant?	Notes
Requirements	No	No PLM/requirements offering
Industrial Design	No	—
Concept Design	Partial	Early feasibility via fast filling simulation; the AI Solver is aimed exactly here
CAD Modelling	No	Imports CAD; does not author it (TopSolid/CAD-integration partnerships)
Engineering Review	Partial	Cadmould Viewer for sharing results
Simulation	Core strength	Cadmould Flex: filling, packing, cooling, warpage, fibre orientation⁵
DFM	Partial — via simulation	Simulation-driven (weld lines, sink marks, warpage); no rule-based DFM checker product
Tool Design	Partial	Simulation feedback to mould layout (gates, cooling); no tool-design CAD
Mould Flow	Core strength	The product's home ground; AI Solver adds a learned fast path
Prototype	Partial	Virtual prototyping by simulation
Validation	Partial	Classical solver positioned as the "final verification" step
Manufacturing Engineering	Partial	Varimos Real optimizes real machine set-up and monitors quality
Production Release	No	—

What makes SIMCON structurally different: a narrow specialist, but the only vendor in this study that today lets the public run a learned neural moulding solver in a browser — and it explicitly pairs the AI with its classical solver ("explore with AI, verify with physics") rather than replacing it. [Synthesis]

AI Strategy¶

Public AI Vision¶

SIMCON's stated vision is a two-speed workflow: the AI Solver as a "high-speed compass" for the iterative design phase, and the classical Cadmould Flex solver as "the definitive map for final validation." CEO Bastiaan Oud: "For decades, the industry has accepted that high-fidelity simulation requires hours of computation... Together, they create an end-to-end workflow that is both incredibly fast and reliable." [Evidence + Marketing] ¹

Investor Statements¶

SIMCON is private; there are no earnings calls. The closest analogue is the Fortino Capital majority investment (June 2024), which framed the thesis as accelerating product innovation and international growth on top of SIMCON's "strong reputation and customer relationships." Twenty-one months later the AI Solver launched — consistent with the investment being, in part, the funding vehicle for the AI push. [Evidence] ⁹ [Inference — the causal link]

Engineering AI Strategy¶

Three distinct mechanisms, which this report deliberately keeps separate: [Synthesis]

Varimos (2010s): automated virtual DoE + fitted meta-model — classical optimization wearing an AI label.²¹ [Evidence + Marketing] ⁷
Varimos Real: the same statistics pointed at measured machine trials — genuine data-fitting, small scale, per mould. [Evidence] ⁸
Cadmould AI Solver (2026): a large learned neural surrogate — genuine modern ML, built with a specialist partner (Emmi AI) rather than in-house alone. [Evidence] ¹

The build-vs-buy pattern mirrors Dassault (which buys LLM capability from Mistral): SIMCON supplied the physics, solver and training corpus; Emmi AI supplied the neural architecture and training expertise. [Synthesis]

Timeline of AI Evolution¶

Date	Event
1988	SIMCON founded — RWTH Aachen spin-off (Dr. Paul Filz)¹⁰⁹
2010s	Varimos launched: automated DoE over Cadmould; later "Varimos AI" branding; Varimos Real for physical trials⁷⁸
Jun 18, 2024	Fortino Capital takes majority stake⁹
Dec 2024	Emmi AI founded in Linz (JKU-Linz ecosystem, CSO Johannes Brandstetter)¹⁶
Feb 2025	AB-UPT architecture paper on arXiv (Emmi AI + Brandstetter)¹⁸
Apr 25, 2025	Emmi AI raises €15M — largest-ever Austrian seed¹⁵
Feb 10, 2026	Emmi AI announces NeuralMould — its injection-moulding Large Engineering Model, built in partnership with SIMCON¹³
Mar 18, 2026	SIMCON launches Cadmould AI Solver — research preview + Early Access Partner Program¹
May 19, 2026	Mistral AI acquires Emmi AI; team joins Mistral, Linz becomes a Mistral office¹⁷

Products Relevant to Engineering¶

Cadmould Flex — the classical simulation product. Simulates filling dynamics, packing (freezing, sink marks), cooling-system thermal behaviour, shrinkage and warpage (deformed geometry output), and fibre orientation, on SIMCON's proprietary 3D-F adaptive mesh technology (a trademarked meshing approach that underpins its speed claims). Sold by subscription ("Flex"). Moulded-parts angle: this is the moulded-parts product. AI: none — physics. [Evidence] ⁵⁶

Varimos / Varimos AI — automation layer on Cadmould. It generates a design of experiments (a structured grid of candidate geometry/process variants), runs the simulations in parallel, fits a meta-model (a response surface — a statistical formula fitted to the simulation results), and suggests optimal settings against the user's targets. Vendor materials call this "artificial intelligence." Under the study's test this is automated search plus curve fitting, recomputed per project — the same category as SIGMASOFT's "Autonomous Optimization" (see report 06), and it takes the same tag here.²¹ [Evidence + Marketing] ⁷¹¹

Varimos Real — the DoE machinery pointed at the real machine: it prescribes a set of physical trial settings, fits a model to the measured part quality, predicts quality for untried settings, and — per vendor material — "can learn how mold and machine sensor data predict part quality" for use in production monitoring. This is genuine statistical learning from measured data, but per-mould and small-sample. [Evidence] ⁸

Cadmould AI Solver — the reason this report exists; detailed in §5. [Evidence] ²

Cadmould Viewer / services — a free results viewer for sharing simulation output with colleagues who do not run the solver, plus outsourced simulation services for teams without in-house capacity. AI: none. [Evidence] ⁴

AI Capabilities¶

5.1 Cadmould AI Solver — the headline capability¶

Description. A transformer-based neural network that predicts injection-moulding filling behaviour — melt-front progression, pressure, temperature and shear-rate fields, plus pressure/flow-rate curves at the gates — directly from part geometry, gate locations, material and process settings, without running the physics equations at answer time. SIMCON calls it "the world's first neural physics engine for plastic injection molding" and (with Emmi AI) "the world's first Large Engineering Model" for the domain — a deliberate echo of "Large Language Model." [Evidence + Marketing] ¹²¹⁴

The honest-AI test: pass. The system unambiguously learns from data. The training corpus is "hundreds of terabytes" — over one million transient simulation trajectories across thousands of materials and systematically varied process conditions — generated by SIMCON's classical Cadmould solvers. This is the real thing, not relabelled optimization. [Evidence] ¹³¹³

...but taught by simulation, not by factories. Note what the training data is: synthetic, solver-generated physics, not measured factory outcomes. The AI Solver learns to imitate Cadmould, so at best it approaches Cadmould's accuracy — it cannot know anything the physics model does not. This confirms the study's cross-vendor finding: even the most advanced learned moulding solver on the market learns from simulations, not from real-world defect/warpage/scrap data. [Evidence] ³ [Synthesis]

Architecture. Emmi AI's model page describes NeuralMould as an AB-UPT variant — "Anchored-Branched Universal Physics Transformers," the architecture Emmi published on arXiv (Feb 2025) for automotive CFD, which handles large meshes by doing neural simulation in a compressed latent space and decoding full-resolution fields on demand. SIMCON's science page matches: a transformer operating in latent space with resolution-agnostic encoding, handling 1M+ node geometries, modelling coupled mass/momentum/energy transport with temperature-dependent viscosity. [Evidence] ¹⁴¹⁸³

Speed claims — a spread, reported honestly. The press release says "up to 1,000×." The product page says ≈200× on mid-tier GPUs and up to 1,000× on high-end systems; the science page says 500–1,000× typical vs. finite-element simulation; Emmi's model page says 200× (1–2 seconds vs. 30–60 minutes). Read "seconds instead of hours" as the robust core of the claim and the specific multiplier as configuration-dependent. [Evidence] ¹²³¹⁴

Accuracy claims — vendor-reported only. Emmi claims "consistent 5% relative errors" vs. traditional solvers; SIMCON's science page says "predominantly single-digit relative errors" with localized double-digit errors in complex regions, validated against production-grade solvers on held-out geometries. SIMCON publishes a downloadable benchmark set (reference geometry + ground-truth solver results) so users can check for themselves. No peer-reviewed paper and no independent third-party validation of the moulding model was found. [Evidence] ¹⁴³ [Evidence of absence]

Stated limitations (from SIMCON itself — unusually candid): [Evidence] ³²

Covers the filling phase only; cooling, shrinkage and warpage are not yet modelled (they are Partner Program roadmap items).
Does not consistently enforce freeze behaviour: melt may be predicted to keep flowing into regions that should have solidified — precisely the flow-hesitation cases that matter in thin-wall automotive parts.
Fine features (small holes, narrow gaps, fine ribs) can be bypassed by the predicted melt front.
Accuracy is bounded by training-regime coverage; positioned for design iteration and sensitivity studies, not high-consequence final validation.

Availability — the crucial caveat. As of August 2026: a free browser-based research preview (fixed geometry set, though with demonstrated generalization to unseen topologies) and an Early Access Partner Program (benchmarking on customer geometries, roadmap influence). General availability has not been announced; there is a waitlist. Anyone describing this as a shipped commercial product is ahead of the evidence. [Evidence] ²

5.2 Varimos AI — automated optimization¶

Automated virtual DoE + meta-model over Cadmould simulations (mechanism in §4). Workflow stage: mould-flow / process optimization. Limitation: per-project search; nothing persists or improves across projects. Classical optimization under an AI label.²¹ [Evidence + Marketing] ⁷

5.3 Varimos Real — learning from physical trials¶

Physical DoE on the moulding machine; fits quality-prediction models to measured trials; vendor materials say it can learn how machine/mould sensor data predict part quality for production monitoring. Why it matters to this study: it is the only capability found across all vendor reports that fits models to measured production data — but it is per-mould statistics, not a transferable learned model. [Evidence] ⁸ [Synthesis]

Engineering Workflow Contribution¶

SIMCON's intended workflow for a moulded automotive part:

import CAD geometry → explore design variants, gate positions and process windows with the AI Solver (seconds per iteration, thousands of variants per day) → narrow to candidates → optimize systematically with Varimos (virtual DoE) → verify the chosen design with the classical Cadmould Flex solver for final sign-off → at the machine, use Varimos Real to set up and stabilize the real process.

The "explore with AI, verify with physics" split is the most architecturally honest AI positioning in this study: the vendor itself tells you not to trust the neural model for final validation. [Evidence] ¹² [Synthesis]

Public Customer Evidence¶

Case Studies¶

Named customers (not AI-specific): Bosch, Continental, Roche and Volkswagen are named in the Fortino investment announcement; the SIMCON site shows logos including Arburg, Diehl, L'Oréal and TE Connectivity. These evidence Cadmould's classical-simulation install base, not AI Solver usage. [Evidence] ⁹⁴
AI Solver customers: none named. The March 2026 launch names no customers; the Partner Program was still recruiting at launch. No customer evidence for the AI Solver exists publicly yet. [Evidence of absence] ¹
Emmi AI's model page claims testing on "1000+ tested real company products" — unverifiable as stated (no names, no methodology). [Evidence — vendor claim, thin]¹⁴

Conference Demonstrations¶

The research preview itself is the demonstration: a public, browser-based, no-installation demo on simcon.ai — a stronger form of evidence than a staged conference video, since anyone can probe it. [Evidence] ² [Synthesis]

White Papers / Testimonials¶

SIMCON's "science" page functions as a technical white paper (methodology, error characterization, limitations, downloadable benchmark). Emmi AI's pages carry a testimonial from SIMCON CEO Bastiaan Oud praising the collaboration. [Evidence] ³¹³

The gap. No public story yet combines the AI Solver + a named customer + a quantified engineering outcome. For classical Cadmould the customer base is verifiable; for the AI layer it is not, because the AI layer is months old. [Evidence of absence]

Technical Architecture (Inferred)¶

Evidence¶

Transformer-based neural network; latent-space simulation; resolution-agnostic encoding; 1M+ node geometries; AB-UPT variant. [Evidence] ³¹⁴
Training data: 1M+ transient trajectories, thousands of materials, "hundreds of terabytes" (Emmi says "almost petabytes"), generated by Cadmould Flex solvers with systematic parameter variation. [Evidence] ¹³¹⁴
Runs in a web browser (server-side GPU inference); speedup figures quoted per GPU tier. [Evidence] ²
AB-UPT itself is published (arXiv 2502.09692) and open-sourced by Emmi AI on GitHub — for automotive aerodynamics; the moulding variant is not open. [Evidence] ¹⁸²⁰

Synthesis¶

The division of labour: SIMCON contributed the physics solver, 35 years of material/process know-how and the compute-generated corpus; Emmi AI contributed the neural architecture (AB-UPT lineage from Johannes Brandstetter's neural-PDE research group) and large-scale training. The product is therefore a two-company system: the "brain" (NeuralMould) originated outside SIMCON.

Inference¶

The training corpus is the durable asset SIMCON controls; the architecture know-how walked into Mistral with the Emmi acquisition. Who owns the trained moulding model weights is not publicly stated — a material open question for SIMCON's roadmap (warpage, shrinkage) if the partnership changes shape.
Because the teacher is Cadmould, systematic errors of Cadmould's physics (material models, 3D-F approximations) are inherited by the student network — an error floor no amount of AI scaling removes.

AI Technologies¶

In one list: neural surrogate simulation (transformer / AB-UPT-variant "Large Engineering Model" for filling — genuine ML)³¹⁴; classical DoE + response-surface optimization (Varimos — AI-labelled, not ML)⁷²¹; statistical process modelling from physical trials (Varimos Real)⁸; no LLMs, no agents, no knowledge graphs, no vision AI anywhere in the offering. [Evidence + Synthesis]

Research Publications¶

Papers¶

AB-UPT (Emmi AI + Brandstetter et al., arXiv 2502.09692, Feb 2025): "Scaling Neural CFD Surrogates for High-Fidelity Automotive Aerodynamics Simulations via Anchored-Branched Universal Physics Transformers" — the published architecture the moulding model descends from. Note: the paper is about car aerodynamics, not moulding. [Evidence] ¹⁸
A follow-up, "AB-UPT for Automotive and Aerospace Applications" (arXiv 2510.15808), extends the line. [Evidence] ¹⁹
No peer-reviewed publication on NeuralMould / the Cadmould AI Solver itself was found. The moulding-specific validation is vendor-published only (the science page + downloadable benchmark). [Evidence of absence] ³

Patents¶

No SIMCON AI patents were identified in this research pass (not exhaustively searched — noted in Appendix C). [Evidence of absence — weak]

Standards¶

No standards-body activity found. [Evidence of absence]

Open Source¶

SIMCON: no open-source AI footprint found. [Evidence of absence]
Emmi AI (the partner): meaningfully open — AB-UPT code on GitHub (Emmi-AI/anchored-branched-universal-physics-transformers) and NeuralDEM released open-source (Sept 2025). The moulding model itself is not open; what is open is the architecture family. [Evidence] ²⁰¹⁶

This is the mirror image of Dassault (§11 of report 01): there the platform giant is closed and the open building blocks come from elsewhere; here the tiny vendor's partner open-sources the architecture while the trained domain model stays proprietary. [Synthesis]

Engineering Service / Platform Mapping¶

Cadmould is a standalone tool, not a platform. It imports geometry from mainstream CAD (CATIA-originated STEP files included) and exports results to structural solvers, metrology software or the moulding machine. Distribution runs through Siemens' Advanced Partner Alliance via Altair Units licensing — so a Siemens/Altair customer can draw Cadmould from the same license pool as Siemens tools — plus regional resellers and a TopSolid CAD partnership. Strategic irony: SIMCON's channel partner Siemens ships its own competing moulding tools, and SIMCON's AI partner now belongs to Mistral, which is also Dassault's AI partner. The small specialist is threaded between giants. [Evidence] ¹¹ [Synthesis]

Engineering Intelligence Stack Mapping¶

Intent — none. No requirements capture; input is geometry + process settings. [Evidence of absence]
Knowledge — implicit: 35 years of material data and solver know-how, now distilled into a training corpus. No queryable knowledge product. [Synthesis]
Reasoning — the strongest layer: physics simulation (classical), DoE search (Varimos), and a learned surrogate (AI Solver) — three reasoning modes over the same domain. [Evidence] ⁵⁷²
Execution — partial: exports set-points toward the machine; Varimos Real closes part of the set-up loop. [Evidence] ⁸
Feedback — the most interesting near-miss in the study: Varimos Real does fit models to measured trials, but per-mould; nothing feeds measured outcomes back into the AI Solver's training. The learning loop from real factories remains open — here as everywhere. [Evidence + Synthesis] ⁸³

Strengths¶

First mover with genuine ML in moulding simulation — a public, learned, transformer-based filling solver, demonstrably ahead of every larger rival covered in this study. [Evidence] ¹ [Synthesis]
Honest engineering culture: published error characterization, published failure modes, downloadable benchmark, explicit "don't use AI for final validation" positioning. Rare and creditable. [Evidence] ³
A real training-data moat: the corpus (10⁶+ trajectories, thousands of materials) required owning a validated solver and material database — exactly the proprietary assets this study's research identifies as the moat. [Evidence] ³ [Synthesis]
Deep, focused domain expertise since 1988; blue-chip classical customer base (VW, Bosch, Continental). [Evidence] ⁹¹⁰
Growth capital (Fortino) and broad distribution (Siemens/Altair channel). [Evidence] ⁹¹¹

Weaknesses¶

The AI Solver is not yet a product — research preview + early access; filling only; no warpage (the automotive money question); no GA date; no named AI customers. [Evidence] ²¹
Dependency on an acquired partner: the neural expertise sits with Emmi AI, now inside Mistral. Continuity of the partnership, and ownership of the trained model, are publicly unresolved. [Evidence] ¹⁷ [Inference]
Known accuracy failure modes (freeze behaviour, fine features) sit exactly where thin-wall automotive parts are hardest. [Evidence] ³
No independent validation — all accuracy numbers are vendor-published. [Evidence of absence] ³
Varimos "AI" branding invites the AI-washing discount on an otherwise honest story.²¹ [Evidence + Marketing] ⁷
Small company (≤50 people) against giants; narrow product surface (no CAD, no PLM, no DFM rules product). [Evidence] ¹⁰ [Synthesis]

Current Gaps (largely manual today)¶

Warpage/shrinkage judgement still requires the classical solver (hours, not seconds). Interpreting results into design changes remains human. No rule-based DFM checker, no geometry-retrieval over past projects, no learning from field quality data into the AI model. The AI Solver's fixed-geometry preview means real customer parts still go through the Partner Program, manually. [Evidence + Synthesis] ²

Future Direction¶

Stated roadmap: shrinkage and warpage prediction for Partner Program members; expanded geometry coverage; accuracy scaling with more training data. [Evidence] ¹³
My read: the strategic race is to reach GA with warpage before Siemens/Altair's physics-AI stack matures in the same direction — the window in which a 50-person company holds a demonstrable lead over the giants is unlikely to stay open long. The Mistral–Emmi question (who trains the next model?) is the biggest single uncertainty. [Inference]

Relevance to Automotive Moulded Parts¶

Capability	Strength	Notes
Plastic Part Design	Indirect	Simulation feedback only; no authoring
Surface Design (Class A)	None	—
CAD Automation	None	Imports CAD; TopSolid/reseller integrations
DFM	Partial — simulation-driven	Weld lines, sink marks, warpage from physics; no rules product; AI Solver makes DFM-by-iteration fast
Tool Design	Partial	Gate/cooling layout evaluation
Mould Flow	Core — plus the only public learned solver	Cadmould Flex (full physics) + AI Solver (filling in seconds, preview)
Manufacturing Engineering	Partial	Varimos Real machine set-up and monitoring
Quality	Partial	Varimos Real quality prediction from sensors (per mould)
Engineering Knowledge Reuse	Weak	Knowledge is in the corpus and the staff, not a queryable product

Critical finding. For the specific question this study asks — is anyone shipping AI that genuinely learns, applied to injection moulding? — SIMCON is the closest thing to a "yes" found anywhere in the research: a real transformer surrogate, publicly demonstrable, honestly documented. The three qualifiers that must always travel with that sentence: it is a preview, not a product; it does filling, not warpage; and it learns from simulations, not from factories. [Synthesis]

Automotive specifics. The claimed sweet spot — exploring thousands of gate/process/design variants per day — maps directly onto automotive Tier-1 workflows (connector housings, clips, interior trim, thick-thin transitions). But the stated failure modes (freeze behaviour in thin walls, bypassed fine ribs) are exactly automotive-critical features, so the vendor's own "verify with the classical solver" caveat is not humility, it is necessity. [Evidence] ³ [Synthesis]

Key Takeaways¶

SIMCON — ~26–50 people, Würselen, founded 1988 out of RWTH Aachen — launched the Cadmould AI Solver on March 18, 2026: a transformer neural surrogate for injection-moulding filling, co-developed with Emmi AI. [Evidence] ¹¹⁰
It passes the study's honest-AI test — it learns from data (10⁶+ simulation trajectories, "hundreds of terabytes"). The most advanced learned moulding solver publicly demonstrated by any vendor in this study. [Evidence] ³ [Synthesis]
Availability caveat: research preview + Early Access Partner Program; filling phase only; warpage/shrinkage are roadmap; GA unannounced; no named AI customers. [Evidence] ²
Accuracy is vendor-reported only (~5% typical relative error, localized double-digit errors; freeze behaviour and fine features are admitted failure modes). No peer review, no independent validation. [Evidence] ³
It is taught by simulation, not by factories — confirming the study's cross-vendor finding that no one learns from measured production outcomes at scale. [Evidence] ³ [Synthesis]
Varimos "AI" is classical DoE + meta-modelling — [Evidence + Marketing], same verdict as SIGMASOFT's Autonomous Optimization.²¹⁷
Varimos Real is the study's closest sighting of learning from real machine data — per-mould statistics, not a learned product. [Evidence] ⁸
Emmi AI was acquired by Mistral (May 19, 2026) two months after the launch; the partnership's future and model ownership are publicly unresolved — the biggest risk to SIMCON's lead. [Evidence] ¹⁷
Classical customer base is blue-chip automotive (VW, Bosch, Continental named by its investor); distribution runs through the Siemens/Altair channel. [Evidence] ⁹¹¹
Net: a small specialist has out-shipped the giants on genuine moulding ML — but "out-shipped" currently means a public preview, and the window to convert it into a product is short. [Synthesis]

References¶

Primary Sources — SIMCON¶

Cadmould AI Solver launch press release, Business Wire, Mar 18 2026 — https://www.businesswire.com/news/home/20260318680159/en/SIMCON-Unveils-Worlds-First-Large-Engineering-Model-for-Plastic-Injection-Moulding (mirrors: Morningstar; Medianet News Hub — https://newshub.medianet.com.au/2026/03/simcon-unveils-worlds-first-large-engineering-model-for-plastic-injection-moulding/144600/)
Cadmould AI Solver product page — https://www.simcon.ai/en-us/solutions/cadmould-ai-solver-injection-molding-simulation
"The Science of AI Simulation in Injection Moulding" — https://www.simcon.ai/en/solutions/cadmould-ai-solver-scientific-research
SIMCON homepage — https://www.simcon.ai/en-us/
Cadmould Flex features (incl. Varimos AI / 3D-F) — https://www.simcon.ai/en-us/solutions/cadmould-plastic-injection-molding-simulation-software/features
Legal notice (registry, managing director) — https://www.simcon.ai/en-us/legal/legal-notice
Varimos (legacy product page, redirects to simcon.ai) — https://www.simcon.com/varimos-injection-molding-doe
Varimos Real (legacy product page) — https://www.simcon.com/varimos-real
Varimos Real webinar page — https://www.simcon.ai/en/learn-and-support/webinars-and-videos/smarter-sampling-varimos-real-optimise-process

Primary Sources — Emmi AI / Mistral¶

NeuralMould announcement, Feb 10 2026 — https://www.emmi.ai/news/neuralmould-our-first-digital-engineer
NeuralMould model page — https://www.emmi.ai/models/neuralmould
€15M seed round, Apr 25 2025 — https://www.emmi.ai/news/emmi-ai-raises-eur-15m-to-bring-ai-to-the-heart-of-industrial-engineering
"Building the Frontier Lab for Industrial Engineering" — https://www.emmi.ai/news/building-the-frontier-lab-for-industrial-engineering
Mistral AI acquires Emmi AI, May 19 2026 — https://www.emmi.ai/news/mistral-ai-acquires-emmi-ai
AB-UPT paper, arXiv 2502.09692 — https://arxiv.org/abs/2502.09692
AB-UPT for Automotive and Aerospace Applications, arXiv 2510.15808 — https://arxiv.org/abs/2510.15808
AB-UPT code — https://github.com/Emmi-AI/anchored-branched-universal-physics-transformers

Corporate / Channel Sources¶

Fortino Capital majority investment, Jun 18 2024 — https://www.fortino.capital/news/fortino-capital-invests-simcon-help-engineers-achieve-better-results
VDWF member directory (founding 1988, 26–50 employees) — https://www.vdwf.de/netzwerk/mitglieder/simconkunststofftechnischesoftwaregmbh.html
Siemens Advanced Partner Alliance: Cadmould & Varimos — https://www.siemens.com/en-us/products/siemens-industry-software-cadmould-and-varimos-by-simcon/
Altair Cadmould/Varimos modules — https://altair.com/cadmould-and-varimos/modules/
KI-Opti-Pack project partner listing — https://ki-hub-kunststoffverpackungen.de/ (SIMCON partner page)

Secondary Sources¶

PlasticsToday, "New Software Slashes Injection Molding Simulation Time" — https://www.plasticstoday.com/injection-molding/large-engineering-model-software-slashes-injection-molding-simulation-time
Design-2-Part Magazine, Jun 5 2026 — https://www.d2pmagazine.com/2026/06/05/large-engineering-model-for-plastic-injection-molding-reduces-computation-times/
MoldMaking Technology product note — https://www.moldmakingtechnology.com/products/simcon-introduces-injection-molding-simulation-tool-with-ai-2
The Machine Maker launch coverage — https://themachinemaker.com/news/simcon-launches-cadmould-ai-solver-redefining-speed-and-efficiency-in-injection-moulding-simulation/
Invest in Austria on Mistral–Emmi — https://investinaustria.at/en/blog/mistral-ai-acquires-austrian-ai-startup-emmi-ai/
FinSMEs on Emmi seed, May 2025 — https://www.finsmes.com/2025/05/emmi-ai-raises-17-1m-in-seed-funding.html
CDG (UK reseller) Varimos page — https://cdg.uk.com/product/varimos/

Appendix A — Timeline¶

1988 — SIMCON founded (RWTH Aachen spin-off, Dr. Paul Filz).¹⁰⁹ → 2010s — Varimos automated DoE; Varimos Real for machine trials.⁷⁸ → Jun 18 2024 — Fortino Capital majority stake.⁹ → Dec 2024 — Emmi AI founded in Linz.¹⁶ → Feb 2025 — AB-UPT on arXiv.¹⁸ → Apr 25 2025 — Emmi's €15M record Austrian seed.¹⁵ → Feb 10 2026 — Emmi announces NeuralMould (with SIMCON).¹³ → Mar 18 2026 — Cadmould AI Solver launched (research preview + Partner Program).¹ → May 19 2026 — Mistral AI acquires Emmi AI.¹⁷

Appendix B — Glossary¶

Cadmould — SIMCON's classical injection-moulding simulation product (filling, packing, cooling, shrinkage, warpage, fibre orientation). "Flex" is the subscription edition.
3D-F — SIMCON's trademarked adaptive meshing technology underlying Cadmould's speed.
Varimos — Cadmould's automation layer: design-of-experiments plus a fitted meta-model, marketed as "Varimos AI."
Design of experiments (DoE) — a structured plan of trial runs (virtual or physical) chosen to map cause and effect efficiently.
Meta-model / response surface — a simple statistical formula fitted to a batch of results, used to predict outcomes for untried settings.
Transformer — the neural-network architecture behind modern language models; here applied to physics fields instead of words.
Neural surrogate — a neural network trained on simulation results that then imitates the simulator at a fraction of the cost.
Large Engineering Model (LEM) — Emmi AI/SIMCON's coined term (echoing "Large Language Model") for a big neural surrogate covering a whole engineering domain.
AB-UPT — Anchored-Branched Universal Physics Transformers: Emmi AI's published architecture for large-mesh neural simulation; NeuralMould is a variant.
NeuralMould — Emmi AI's name for the injection-moulding model inside the Cadmould AI Solver.
Research preview — a free, public, limited demo of an unreleased product; not general availability.
Freeze / flow hesitation — melt solidifying in thin sections and stalling the flow front; a known weak spot of the AI Solver's predictions.
Filling / packing / warpage — the three headline stages of moulding simulation: how the cavity fills; how pressure compensates shrinkage; how the part distorts after ejection.

Appendix C — Notes (thinnest-evidence areas to revisit)¶

All accuracy figures are vendor-published (SIMCON science page, Emmi model page). No peer-reviewed paper on NeuralMould, no independent benchmark. The downloadable benchmark set enables — but does not constitute — third-party validation.
Training-corpus size wording varies: "hundreds of terabytes" (SIMCON press release) vs. "almost petabytes" (Emmi model page). Treat as order-of-magnitude marketing, not audited fact.
Post-acquisition status of the SIMCON–Emmi partnership and ownership of the trained moulding model: no public statement found after Mistral's May 19, 2026 acquisition. This report's §15/§17 risk framing is inference. Re-checked 2026-08-02: still unresolved. SIMCON's live AI Solver page (fetched this date) still credits Emmi AI as partner and mentions Mistral nowhere; Mistral/Emmi's acquisition announcement says nothing about partner projects; no press, LinkedIn or trade item May–Aug 2026 clarifies continuity. GA status unchanged: still a "research preview" limited to the filling phase ("does not yet model cooling or warpage"), free browser demo + waitlist + Partner Program. Warpage roadmap: trade coverage (PlasticsToday; Design-2-Part, Jun 5 2026) says shrinkage/warpage extension is under development, Partner Program members get early access — no dates. Item stays OPEN.
Employee count (26–50) is from the VDWF member directory; possibly stale post-Fortino. SIMCON publishes no headcount or revenue.
"Best-selling German injection-moulding simulation software" is a directory/vendor formulation, not independently audited.
Customer names (VW, Bosch, Continental, Roche) come from the investor's announcement and website logos — classical-Cadmould evidence; none is tied to the AI Solver, and Emmi's "1000+ tested real company products" is unverifiable as stated.
Varimos mechanism detail (meta-model fitting method) is drawn from legacy vendor pages and reseller descriptions; the exact fitting algorithm (polynomial, kriging, neural) is not publicly specified.
Patent search was not exhaustive; "no SIMCON AI patents" is a weak negative.
The original Business Wire URL timed out during research; content was verified via the Medianet mirror and trade coverage of the same release.

Executive Summary¶

Who they are. SIMCON is a small, privately held German software company in Würselen near Aachen, founded in 1988 as a spin-off from RWTH Aachen University by Dr. Paul Filz. It employs roughly 26–50 people (directory-derived — the company publishes no headcount). Its product, Cadmould, simulates the injection-moulding process (filling, packing, cooling, shrinkage, warpage, fibre orientation). Since June 2024 the growth investor Fortino Capital holds a majority stake, and Dr. Bastiaan Oud is CEO. [Evidence] ¹⁰⁹⁶
Why this report exists. On March 18, 2026 SIMCON launched the Cadmould AI Solver, billed as the "World's First Large Engineering Model for Plastic Injection Moulding" — a transformer neural network (the architecture family behind modern language models, applied here to physics fields), co-developed with the Austrian startup Emmi AI, trained on simulation data, and claimed to be up to 1,000× faster than classical solvers. [Evidence] ¹
The central honest finding — it is real ML, with a big availability caveat. Applying the study's test ("does the system learn from data, or compute from equations?"), the Cadmould AI Solver passes: it is a genuinely learned neural surrogate, trained on over a million transient simulation trajectories ("hundreds of terabytes") generated by SIMCON's own classical solver. But it is not yet a generally available product. As of this writing it exists as a free browser-based "research preview" (filling phase only, a fixed set of geometries) plus an Early Access Partner Program; shrinkage and warpage prediction are roadmap items, and general availability has not been announced. "First to put a learned moulding solver in public hands" is accurate; "first shipping product" is not yet. [Evidence] ²³¹
Unusually honest about limits. SIMCON's own science page reports "predominantly single-digit relative errors" with localized double-digit errors, and openly lists failure modes: the model does not consistently enforce freeze behaviour (melt can flow into regions that should have solidified), and fine features (small holes, thin ribs) can be bypassed. No peer-reviewed paper or independent validation of the moulding model was found. [Evidence] ³
The rest of the AI story is classical. Varimos, SIMCON's long-standing "AI" optimization product, is automated design-of-experiments (DoE) plus a fitted response-surface "meta model" — automated search by simulation, not a learning system in the modern sense. It gets the study's [Evidence + Marketing] treatment.²¹ One genuine nuance: Varimos Real fits its statistical model to measured machine trials, the closest any vendor in this study comes to learning from real factory data — though per-mould and small-sample, not a persistent learned model. [Evidence] ⁷⁸
A strategic wildcard. Emmi AI — SIMCON's AI co-developer — was acquired by Mistral AI in May 2026, two months after the joint launch. What happens to the partnership and to the underlying model (Emmi's "NeuralMould") under Mistral ownership is publicly unanswered. [Evidence] ¹⁷
Bottom line for the study. A ~50-person German specialist has publicly demonstrated the most advanced learned injection-moulding solver of any vendor examined in this study — ahead of Autodesk (research demo) and Siemens/Altair (physics AI arriving mid-2026) — while being refreshingly candid about its limits. The claim to watch is not the 1,000× speed; it is whether the accuracy, scope (warpage) and availability mature into a real product. [Synthesis]

SIMCON, "SIMCON Unveils World's First Large Engineering Model for Plastic Injection Moulding," Business Wire, Mar 18 2026 — https://www.businesswire.com/news/home/20260318680159/en/SIMCON-Unveils-Worlds-First-Large-Engineering-Model-for-Plastic-Injection-Moulding (verified via mirror: https://newshub.medianet.com.au/2026/03/simcon-unveils-worlds-first-large-engineering-model-for-plastic-injection-moulding/144600/) ↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩
Cadmould AI Solver product page (research preview, Partner Program, speed tiers, filling-only scope) — https://www.simcon.ai/en-us/solutions/cadmould-ai-solver-injection-molding-simulation ↩↩↩↩↩↩↩↩↩↩↩↩↩
SIMCON, "The Science of AI Simulation in Injection Moulding" (architecture, training corpus, error characterization, stated limitations, benchmark downloads) — https://www.simcon.ai/en/solutions/cadmould-ai-solver-scientific-research ↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩↩
SIMCON homepage (products, customer logos, 12,000+ projects, 35+ years) — https://www.simcon.ai/en-us/ ↩↩↩↩
Cadmould Flex features page (filling/packing/warpage/fibre, 3D-F, Varimos AI) — https://www.simcon.ai/en-us/solutions/cadmould-plastic-injection-molding-simulation-software/features ↩↩↩
SIMCON legal notice (entity, HRB 3996 Aachen, Dr. Bastiaan Oud, trademarks) — https://www.simcon.ai/en-us/legal/legal-notice ↩↩↩
Varimos product description (automated DoE, meta-model, "artificial intelligence" wording) — https://www.simcon.com/varimos-injection-molding-doe; Siemens listing ("AI-assisted design variant analysis") — https://www.siemens.com/en-us/products/siemens-industry-software-cadmould-and-varimos-by-simcon/ ↩↩↩↩↩↩↩↩↩↩
Varimos Real (physical DoE, quality prediction from measured trials and sensor data) — https://www.simcon.com/varimos-real; webinar page — https://www.simcon.ai/en/learn-and-support/webinars-and-videos/smarter-sampling-varimos-real-optimise-process ↩↩↩↩↩↩↩↩↩↩
Fortino Capital, "Fortino Capital invests in SIMCON," Jun 18 2024 (majority stake; founder Dr. Paul Filz; customers Bosch, Continental, Roche, Volkswagen) — https://www.fortino.capital/news/fortino-capital-invests-simcon-help-engineers-achieve-better-results ↩↩↩↩↩↩↩↩↩↩↩↩↩
VDWF member directory: SIMCON (founded 1988, Würselen, 26–50 employees) — https://www.vdwf.de/netzwerk/mitglieder/simconkunststofftechnischesoftwaregmbh.html ↩↩↩↩↩↩↩↩↩
Siemens Advanced Partner Alliance: "CADMOULD and VARIMOS by SIMCON" (Altair Units licensing) — https://www.siemens.com/en-us/products/siemens-industry-software-cadmould-and-varimos-by-simcon/ ↩↩↩↩↩
KI-Opti-Pack partner listing, SIMCON — https://ki-hub-kunststoffverpackungen.de/ (partner-organisation page surfaced in search; project-level detail not further verified) ↩
Emmi AI, "NeuralMould: Our First Digital Engineer for Injection Moulding," Feb 10 2026 (1M+ trajectories, SIMCON partnership, Oud quote) — https://www.emmi.ai/news/neuralmould-our-first-digital-engineer ↩↩↩↩
Emmi AI NeuralMould model page (AB-UPT variant, "consistent 5% relative errors," 200×, 1–2 s, "1000+ tested real company products") — https://www.emmi.ai/models/neuralmould ↩↩↩↩↩↩↩↩
Emmi AI €15M seed announcement, Apr 25 2025 (3VC, Speedinvest, Serena, PUSH; largest Austrian seed) — https://www.emmi.ai/news/emmi-ai-raises-eur-15m-to-bring-ai-to-the-heart-of-industrial-engineering ↩↩
Emmi AI, "Building the Frontier Lab for Industrial Engineering" (founded Dec 2024; Brandstetter CSO; co-founders Dennis Just, Miks Mikelsons; NeuralDEM open source) — https://www.emmi.ai/news/building-the-frontier-lab-for-industrial-engineering ↩↩↩
Emmi AI, "Mistral AI Acquires Emmi AI," May 19 2026 (team joins Mistral; Linz becomes Mistral office) — https://www.emmi.ai/news/mistral-ai-acquires-emmi-ai ↩↩↩↩↩
Alkin, Bleeker, Kurle, Kronlachner, Sonnleitner, Dorfer, Brandstetter, "AB-UPT: Scaling Neural CFD Surrogates for High-Fidelity Automotive Aerodynamics Simulations via Anchored-Branched Universal Physics Transformers," arXiv 2502.09692 — https://arxiv.org/abs/2502.09692 ↩↩↩↩↩
"AB-UPT for Automotive and Aerospace Applications," arXiv 2510.15808 — https://arxiv.org/abs/2510.15808 ↩
Emmi AI AB-UPT repository — https://github.com/Emmi-AI/anchored-branched-universal-physics-transformers ↩↩
Why DoE/meta-model optimization is not AI in the modern sense — see the concept note What is Optimization technology: virtual design-of-experiments runs many physics simulations across candidate settings and fits a response surface to pick the best; it is automated search, deterministic, and carries nothing between projects — classical optimization, not a learning system. ↩↩↩↩↩↩↩