nTop and Neural Concept: Engineering the Next Generation of AI-Driven Product Design

Company Profiles

nTop (formerly nTopology) was founded by Brad Rothenberg in 2015 with a specific thesis: that the reason engineering simulation and manufacturing feedback loops are so slow is not the simulation software — it is the CAD geometry representation. Traditional B-rep (boundary representation) solid modeling, the geometry format underlying every major CAD system, is efficient for drafting but hostile to optimization. Changing a topology-optimized geometry in a B-rep model is difficult. Lattice structures are nearly impossible to represent accurately. nTop replaced the representation with a field-based approach — geometry as a mathematical field function rather than a collection of faces and edges — that is natively compatible with optimization algorithms.

Neural Concept was founded in Switzerland by Thomas von Tschammer and team, emerging from the ETH Zurich-adjacent startup ecosystem. The company applies deep learning to a specific problem: predicting the results of finite element analysis (FEA) and computational fluid dynamics (CFD) simulation without running the full simulation. In late 2024, Neural Concept raised $100 million from Goldman Sachs — one of the largest funding rounds in manufacturing AI — to accelerate deployment across aerospace and automotive customers.

---

The Engineering Design Bottleneck

The classical engineering design process has a fundamental throughput constraint: the simulation-evaluation loop.

An engineer designs a part. To know if the design will work — structurally, thermally, aerodynamically — they run a simulation (FEA, CFD, or both). For a complex aerospace part, a single FEA run on a high-fidelity mesh takes hours to days on an HPC cluster. If the simulation reveals a problem, the engineer modifies the design and runs it again. In a typical complex component development program, this loop runs dozens to hundreds of times before a design is accepted.

The result: the design space that engineers actually explore is a tiny fraction of what is physically possible. Not because engineers lack creativity or capability — because they cannot afford the compute time to evaluate more candidates.

This is the problem both nTop and Neural Concept are targeting from different angles.

---

What nTop Built

The Geometry Representation Problem

Rothenberg's insight was that B-rep geometry — the foundation of CATIA, NX, Creo, and SolidWorks — was designed for drafting and manufacturing documentation, not optimization. B-rep models define geometry as collections of faces, edges, and vertices. This works well for representing the specific geometry an engineer drew. It works poorly for representing the space of geometries an engineer might want to consider.

Topology optimization — finding the optimal material distribution within a design space — produces geometries with organic, non-uniform structures that B-rep cannot represent without massive polygon counts. Lattice structures — repeating micro-structural patterns used in additive manufacturing to achieve high strength-to-weight ratios — are effectively impossible to represent accurately in B-rep.

nTop's field-based representation describes geometry as an implicit function over space. The geometry is not a collection of faces — it is a function that returns "inside material" or "outside material" for any point in 3D space. Optimization algorithms operate natively on this representation. Adding a hole, a lattice infill, or a topology-optimized structure are single function calls, not complex geometry editing operations.

The manufacturing connection is equally important: nTop encodes manufacturing constraints — minimum feature size for additive manufacturing, draft angles for casting, wall thickness for CNC — as additional field functions that modify the design space. The result: designs that come out of nTop optimization are manufacturable by construction, not by inspection after the fact. This eliminates a major category of design-manufacturing iteration: the "we can't actually build this" rework cycle.

Enterprise Adoption Through Education

nTop's deployment model is distinctive: the company invests heavily in customer education, running structured boot camps that teach engineering teams not just how to use nTop but how to redesign their workflows around computational design. This is deliberate. The technology requires a workflow change, not just a tool swap. Engineers who add nTop to a legacy design process get marginal value. Engineers who redesign their process around nTop's capabilities get order-of-magnitude results.

Customers span aerospace (structural brackets, heat exchangers, ducting), automotive (suspension components, structural elements), medical devices (orthopedic implants, bone scaffolds), and advanced manufacturing broadly. The common thread: parts where weight matters, or where additive manufacturing opens geometric freedom that traditional design cannot take advantage of.

---

What Neural Concept Built

Neural Concept's approach is to learn the mapping from geometry to simulation result, rather than computing it from physics each time.

FEA and CFD are expensive because they discretize a geometry into millions of small elements and solve coupled partial differential equations across the entire mesh. The physics is correct. The computation is slow. Neural Concept trains deep learning models on large libraries of FEA/CFD results for families of geometries, and the trained model can predict the simulation outcome for a new geometry in seconds — without running the full simulation.

The accuracy limitation is important to understand: Neural Concept's model is accurate for geometries similar to its training set. For genuinely novel geometries outside the training distribution, accuracy degrades. The intended workflow is not to replace high-fidelity simulation for design sign-off — it is to replace simulation during early-stage design exploration, where engineers are evaluating many candidates quickly and don't need sign-off-level accuracy.

The practical result: an engineer who previously could afford to evaluate 10–20 designs during a design exploration phase can now evaluate thousands. The design space that gets explored expands by two to three orders of magnitude. Better designs get found — designs that would never have been discovered if the engineer had to run a full FEA on each candidate.

The $100M Goldman Sachs Bet

The funding scale signals where the market is going. Goldman Sachs, not typically a manufacturing technology investor, committed $100 million to Neural Concept in late 2024. The rationale: simulation acceleration is a bottleneck across the entire manufacturing value chain — aerospace, automotive, medical devices, consumer electronics, industrial equipment. Every complex product that requires simulation before manufacturing benefits from this technology. The market is enormous, and Neural Concept has technical lead.

The deployment implications: the company is scaling from pilot programs to systematic enterprise deployment across multiple industries simultaneously. At that scale, ROI documentation becomes critical, and Neural Concept is building the case study library to support it.

---

Results

nTop customer outcomes (aerospace):

• Structural bracket designs achieving 30–50% weight reduction versus conventionally designed parts, manufacturable by additive manufacturing
• Design cycle compression from 6–8 weeks (design → simulation → manufacturing review → redesign) to 2–3 weeks, by embedding manufacturing constraints in the design tool
• Heat exchanger designs with 2–3x thermal performance improvement over conventionally designed parts, achieved through topology-optimized channel geometries impossible to design in B-rep CAD

Neural Concept deployment outcomes:

• FEA prediction time for automotive body panel aerodynamics: from 8–12 hours (full CFD) to under 2 minutes (Neural Concept prediction) with accuracy sufficient for design screening
• Early-stage design exploration: customers report evaluating 10–50x more design candidates per development cycle
• Time-to-concept-freeze reduction of 30–40% in programs where Neural Concept replaced manual design iteration

---

Lessons Learned

1. The bottleneck is not creativity — it is evaluation. Engineers are not running out of design ideas. They are running out of time to evaluate the ideas they have. Both nTop and Neural Concept remove evaluation time as the constraint.

2. Workflow redesign is mandatory. Both companies have found that customers who try to add their technology to existing workflows get a fraction of the value of customers who redesign around the technology. Boot camps and structured onboarding are investments in workflow transformation, not just product training.

3. Accuracy for the right phase matters more than absolute accuracy. Neural Concept's surrogate models are not as accurate as full FEA. They are accurate enough for early-stage exploration, which is where the simulation bottleneck is most damaging. Matching the accuracy level to the design phase unlocks value that "not as accurate as full FEA" obscures.

4. Manufacturability-in-design eliminates a rework category. nTop's approach of encoding manufacturing constraints in the design environment eliminates the "can we build this?" review cycle that follows design-simulation iterations. This is a category of rework, not an optimization — removing it changes the economics significantly.

5. The capitalization signal matters. $100M from Goldman Sachs for Neural Concept is not just funding. It is a signal that institutional capital with long time horizons sees simulation acceleration as a durable industrial technology, not a startup bet.

---

Implementation Advice

For engineering organizations evaluating design AI: the right entry point depends on your primary constraint.

If your constraint is design quality (weight, performance, cost) and you have additive manufacturing capability — evaluate nTop. The combination of computational design freedom and additive manufacturing unlocks parts that conventional CAD-then-simulate workflows simply cannot produce.

If your constraint is design cycle time and you run large simulation programs — evaluate Neural Concept. The investment in building a training library for your specific geometry families pays off in design cycles measured in days instead of weeks.

Both technologies require investment beyond the license fee: training data for Neural Concept, workflow redesign for nTop. Organizations that treat them as plug-in tools get marginal results. Organizations that treat them as workflow transformation tools get step-change results.

---

About the Source

This case study is drawn from AI Across the Product Lifecycle Episode 25, a podcast conversation with Brad Rothenberg (CEO, nTop) and Thomas von Tschammer (Neural Concept). See also: [[Topology Optimization]], [[Digital Twin in Manufacturing]], [[Additive Manufacturing PLM]], [[Simulation and PLM]].