1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering

Episode Summary

Executive Summary: The episode argues that engineering is entering an AI-accelerated phase analogous to the digitization of software: specialist physics models, paired with agentic workflows and CAD/simulation tools, can shrink iteration cycles from days to minutes and unlock far larger design spaces. Neural Concept is positioning itself as the application layer for engineering AI, starting with per-customer models and moving toward foundation models, while keeping human engineers in the loop for trade-offs, manufacturing constraints, and final decisions.

Main Topics: AI for physics-based engineering (Priority: 5/5): Thomas von Schommer explains how Neural Concept uses specialist models to approximate expensive physics solvers for aerodynamics, crash safety, heat dissipation, and other engineering domains, dramatically speeding up design iteration. From CAD and simulation to agentic workflows (Priority: 5/5): The discussion traces the evolution from manual prototyping to CAD/FEA and now to AI copilots that can interact with CAD, run simulations, and orchestrate optimization loops with engineers. Customer-specific models vs foundation models (Priority: 4/5): Neural Concept currently fine-tunes models on company-specific simulation and test data, but expects domain foundation models—starting with aerodynamics—to emerge and eventually plug into broader engineering workflows. Human-in-the-loop trade-offs and manufacturing constraints (Priority: 5/5): The guest emphasizes that engineering is multi-objective and never fully black-boxed; engineers must retain final judgment because real products involve safety, cost, manufacturability, and organizational IP. Competitive dynamics in automotive and F1 (Priority: 4/5): Public examples like Jaguar Land Rover and Formula One show major speedups and illustrate how compute limits, workflow automation, and engineering agility can create or widen competitive gaps. Industrial disruption and organizational change (Priority: 4/5): The conversation highlights that legacy OEMs may struggle to adopt AI workflows as fast as digital-native manufacturers, creating opportunities for new entrants and accelerating competitive divergence.

Key Arguments: Engineering AI follows a recurring pattern seen in proteins and other domains: digitize a physical process, learn from data, and replace slow solvers with much faster specialist models. AI does not replace engineering judgment; it compresses the search space so experts can evaluate many more viable options and focus on higher-level trade-offs. Neural Concept’s models can be trained on both simulation and real test data, which helps capture phenomena traditional solvers miss and preserves company know-how. The company is moving from models that validate designs toward agentic systems that can also generate and modify geometry in CAD. Foundation models for engineering are likely, but near-term value comes from combining general-purpose reasoners with specialized physics tools and domain constraints. Manufacturing constraints and cost must be embedded early in optimization, or design improvements will be unusable in production. AI adoption will create widening performance gaps between companies that integrate it deeply and those that remain on old workflows. Formula One is a useful proving ground because compute limits force teams to maximize workflow efficiency and expose the value of better AI tools. Autonomous-vehicle and EV trends may intensify design differentiation, but only if companies encode their proprietary engineering practices into AI workflows. The biggest near-term bottleneck is less model capability than infrastructure, governance, and data flow integration across large organizations.

Data Points: Jaguar Land Rover aerodynamic evaluations: from 50 designs/day to 1,500 designs/day - Publicly shared example of AI-driven aerodynamic workflows using Neural Concept Battery cold plate development cycle: 80% faster - Supplier example cited for thermal-management design acceleration Battery cooling performance improvement: 20% better cooling - Example of better-performing design enabled by larger search space Aerodynamic improvement: 2% to 5% more aerodynamic - Illustrative gains from exploring more options OEM product development cycle: 40 to 60 months - Typical U.S./Western European car development timeline mentioned in contrast to China China automotive development cycle: 18 to 24 months - Faster iteration and manufacturing cadence cited as competitive benchmark Flagship discipline speedup target: 20% to 40% - Year-one AI-led workflows for crash safety, aerodynamics, and powertrain Cross-discipline cycle reduction target: 50% to 60% - Year-two orchestration across crash, aero, thermal, manufacturing, and constraints F1 compute limit: CPU-hours capped per team - Formula One teams are restricted in aerodynamic simulation compute Neural Concept founding: 2019 - Guest says the company began with vision-based models for geometry-to-physics prediction Engineering breadth: thousands of engineers - Large OEMs simulate many components and subassemblies across multiple physics domains Podcast scale reference: 350 episodes - Host notes prior lack of coverage on AI for engineering on the show

Pivotal Quotes: "“We are not in the world where we're just sending a spec sheet to the AI and expect the AI to get back to us as a black box with the best optimized car.”" — Thomas von Schommer: Explaining why engineering AI remains human-in-the-loop and trade-off driven "“If you can get results in minutes, it means that you don't explore 50 designs a year... but now thousands of different options.”" — Thomas von Schommer: Describing the core productivity leap from AI surrogate models over physics solvers "“The winning companies are the ones that are going to be able to code the best practice brand within this AI workflows.”" — Thomas von Schommer: On competitive advantage and the role of organizational know-how in AI systems

Implications: Engineering is becoming AI-native: faster simulation, broader search, and agentic CAD workflows will compress product cycles, widen industry gaps, and reward firms that encode manufacturing know-how and governance into AI systems while keeping humans on final trade-offs.