What's Actually Changing in CAD, FEA, and Engineering Simulation

Engineering software has developed for decades as a set of separate applications with separate skills attached to them. CAD, meshing, FEA, optimization, and PLM each improved on their own terms, and the effort of moving work between them was mostly absorbed by the engineer. A significant share of any analyst's or designer's week still goes into that transfer: cleaning up geometry for a solver, rebuilding an optimized shape into something editable, re-entering data that already existed somewhere else in the project.

Over the past year, a fairly consistent pattern has emerged across vendor roadmaps, academic research, and industry bodies like NAFEMS. AI is being applied unevenly across this pipeline, and the areas seeing real progress are not the ones that get the most attention. This piece looks at where the evidence actually points: what's working, what's still experimental, and where the harder problems remain unsolved.

AI in CAD is augmenting existing workflows, not replacing them

The AI features that have reached production in mainstream CAD tools are narrow by design: sketch constraint suggestions, automatic drawing generation, fastener and component recognition, command guidance. Autodesk has described its own roadmap as a progression from task-level automation toward workflow automation, with system-level autonomy positioned as a longer-term goal rather than a current capability. Siemens and Dassault are following a similar path with their own assistants.

Generative CAD, meaning models that produce a design from a prompt, image, or point cloud, remains largely in research. Recent work such as DreamCAD and CADFit has made progress on generating editable B-rep and parametric surface models rather than static meshes, which is the correct target, since a model that only looks like a part is far less useful than one that behaves like one in a feature tree.

The gap between the two is still substantial. A study published in May 2026 tasked AI agents with producing fully assembled STEP files that had to pass FEA-based validation checks. None of the tested configurations produced a strict passing result in the main evaluation, and the best-performing setup satisfied roughly 20 percent of the typed requirements on average. The finding is a useful corrective: FEA feedback measurably improves AI-generated CAD, but it does not yet make the output reliable enough to use without engineering review.

The CAD-to-FEA handoff is the more immediate automation target

The step between a CAD assembly and a usable analysis model consumes more engineering time than the solver run itself. Production geometry typically contains details that are appropriate for manufacturing but not for simulation: small fillets, logos, fastener threads, gaps and overlaps between mating parts, thin sections that need midsurface extraction, and joints that need to be idealized as beams, connectors, or contact pairs. Every design revision after that requires the same idealization to be redone, often manually, with loads and boundary conditions reattached by hand.

A 2025 survey of AI methods for geometry preparation and meshing frames the current state accurately: AI is functioning as an assistive layer over conventional geometric and numerical methods, not a replacement for them. Solvers still require valid elements, consistent topology, and predictable convergence behavior, and that requirement is not something a language model changes.

What is being built toward is closer to a semantic CAD-to-CAE compiler: software that recognizes engineering features and interfaces, selects the appropriate analysis abstraction, preserves named regions and design intent through a revision, generates contacts and mesh controls, and flags setup decisions that look inconsistent with prior practice. Compared with general text-to-CAD generation, this is a narrower and more tractable problem, and it addresses friction that every analysis team already deals with on a recurring basis.

FEA is splitting into a high-fidelity tier and a predictive tier

A two-tier structure is forming in simulation software.

The first tier is traditional FEA and CFD, which remains the source of authoritative results and the training data that everything downstream depends on. This tier is being accelerated through GPU compute, cloud HPC, and improved parallel solvers, with NVIDIA and the major CAE vendors actively optimizing for GPU hardware.

The second tier consists of surrogate models trained to map geometry, materials, and boundary conditions directly to fields such as stress, displacement, temperature, or flow. Ansys SimAI, Siemens Simcenter PhysicsAI, and NVIDIA PhysicsNeMo all represent this direction, using geometric deep learning and neural operator architectures trained on historical simulation results.

The value proposition is straightforward: fast screening across many design variants, optimization loops that can afford thousands of evaluations instead of a handful, and earlier feedback for a designer before a specialist runs the final validated case. The constraint is equally straightforward. A surrogate model is reliable within the envelope of geometries, loads, and behaviors it was trained on, and its performance outside that envelope is not something the model can be counted on to flag for itself. NAFEMS lists robustness, data generation, generalization, and validation as the open questions in this area, and none of them have a settled answer yet.

Topology optimization still produces a shape, not a manufacturable part

Topology optimization reliably produces high-performing structural shapes, but the output is typically an organic mesh that requires substantial manual rework before it can be dimensioned, edited, or manufactured. That rework step offsets a meaningful portion of the time savings the optimization was meant to deliver.

Current research is addressing this from several angles: automatic reconstruction of topology-optimization output into CAD, optimizing parameters directly within an existing feature tree, differentiating through NURBS surfaces to keep the result smooth and editable, and incorporating manufacturing constraints into the optimization loop itself rather than applying them afterward. A related and more experimental direction is the differentiable engineering pipeline, where gradients are propagated from a performance metric back to geometry, in some cases by replacing non-differentiable meshing or solving steps with differentiable surrogates.

The target these approaches are working toward is a system that can take a stated objective, such as minimizing mass subject to stiffness, fatigue, thermal, and machining constraints, and return a clean parametric model rather than a mesh that has to be reconstructed by hand.

Agentic workflows are handling coordination, not judgment

Engineering agents currently in use or in pilot are concentrated on operational tasks: selecting an approved simulation template, submitting solver runs to HPC, managing parameter sweeps for design-of-experiments studies, monitoring for failed jobs, extracting results, and assembling reports. NAFEMS' 2026 CAE-AI program reflects this same scope, covering practical agentic workflows, CAD-change impact prediction, and AI-assisted optimization alongside more established topics.

What is consistently absent from current deployments is agent authority over the physics itself. No agent is approving a nonlinear structural result without a human reviewer. That division makes sense given the asymmetry in cost: an incorrectly scheduled job is recoverable, while an incorrectly approved analysis result is not. At present, the agent's role is closer to an operations coordinator than an analyst, automating execution and information movement while escalating uncertain decisions.

Platform consolidation is being driven by the digital thread

The competitive positioning among major CAD, CAE, and PLM vendors is increasingly about connectivity rather than any single feature. Autodesk Fusion combines CAD, CAM, CAE, electronics, and data management in one environment. Siemens positions NX, Simcenter, and Xcelerator as a connected design-to-manufacturing platform. Onshape links cloud-native CAD and PDM to Arena's PLM and BOM data. Dassault is integrating virtual twins and lifecycle management into 3DEXPERIENCE. Consolidation is also happening structurally: Siemens completed its acquisition of Altair in March 2025, and Synopsys completed its acquisition of Ansys in July 2025.

The underlying objective is to preserve the relationship between a requirement, a geometry revision, a material and manufacturing decision, a simulation assumption, a solver setting, a piece of test evidence, and a released configuration. Without that chain of provenance intact, both AI-generated predictions and automated decisions become difficult to audit after the fact, which is a requirement in most regulated engineering domains regardless of how the analysis was produced.

Validation has become the central constraint

Across CAD generation, surrogate physics, and optimization, the question that industry bodies and vendors are converging on is not whether an AI system can produce a plausible result, but where that result is valid, how uncertain it is, and whether a decision can be certified on top of it. NAFEMS' emphasis on explainability, confidence estimation, and verification and validation for AI-assisted CAE reflects this directly.

This is the practical bottleneck at the moment. The capability to generate a plausible-looking output, whether that's a CAD model, a stress field, or an optimized shape, is advancing quickly. The infrastructure needed to know when that output can be trusted, and to demonstrate that to a reviewer or a regulator afterward, is further behind and is where most of the unresolved engineering work sits.

Where DeepMechanix fits into this

Our work is focused on a narrower slice of this landscape: engineering, procurement, and construction, and pressure vessel design specifically.

EPC work is, at its core, a documentation and calculation pipeline. A project moves through feasibility, engineering, procurement, construction, and operations, and at most stages an engineer is applying codified rules and producing paperwork that has already been produced, in similar form, on other projects governed by the same code. This is the kind of calculation-heavy, standards-bound work identified above as a near-term opportunity: not a general-purpose AI engineer, but a system that owns one domain's rules and failure modes precisely enough to be trusted with them.

DeepMechanix is built around pressure vessel design to ASME Boiler and Pressure Vessel Code, Section VIII. The platform runs the relevant clause checks in a single pass, covering areas such as shell thickness under UG-27, nozzle reinforcement under UG-37, and MDMT determination under UCS-66, and returns each result tied to its governing formula, substituted values, and the specific code paragraph it comes from. This is closer in structure to the CAD-to-CAE compiler and credibility-layer opportunities described above than to a surrogate model: the output is deterministic and auditable against a published standard, and its validity envelope is the code itself rather than a training distribution.

The agentic component follows the pattern seen elsewhere in this space, where agents are given responsibility for execution and documentation before they are given responsibility for engineering judgment. DeepMechanix runs browser-native alongside the CAD, FEA, and plant design tools a team already uses, and it is intended to handle sizing, code checking, and report generation rather than replace the engineering sign-off that follows. Output is exported as a portable, self-contained deliverable that fits into existing document control processes rather than requiring a new one.

Summary

The near-term progress in this field is concentrated at the boundaries between tools rather than inside any single one of them: the handoff between CAD and analysis, the reuse of simulation knowledge a company already has, and the layer of validation and traceability that lets an organization actually rely on an AI-produced result. General-purpose generative design and fully autonomous analysis remain further out, constrained less by model capability than by the difficulty of proving, after the fact, that a given output can be trusted.

That is the specific problem DeepMechanix is built to address within EPC and pressure vessel engineering. If you want to see what a fully clause-cited, code-checked vessel report looks like in practice, you can review the pressure vessel product or join the waitlist for the broader platform.