Generative Engineering
Abstract
Leveraging AI and modern compute to rebuild simulation workflows for mechanical and fluid dynamic engineering, resulting in drastically faster design throughput, iteration cycle and lower cost.
Description
For the past two years I’ve been actively hunting for investments within the engineering simulation workflow, most notably in computational fluid dynamics (CFD), but also in several mechanistic disciplines within robotics. For decades these fields have been dominated by extremely robust, mathematically defined solvers that prove out performance properties (aerodynamics, shear pressure, thermal exchange, etc.) with near certainty. New capabilities and cloud-based collaborative design have been folded into incumbent systems, but they remain largely static and now reflect a Frankenstein’s monster of acquisitions. Because these systems are deterministic, the computational load that any run requires is enormous. It is not uncommon to hear engineers claim that a single simulation can run for a full week, and that’s hoping nothing crashes midway through. This inherent inefficiency, compounded across multiple sub-disciplines, leads to huge amounts of idle time and months-long development cycles that grind the entire innovation process to a crawl.
Mechanical engineering, like so many physical disciplines, has seen a talent flight over the past two decades, which leaves the West in a precarious labor shortage at the worst possible time. Globally there are approximately 325,000 specialist engineers in this field, but nearly 50% are over 50 and looking down the barrel of retirement. By some estimates, we’re currently 18% below where the market needs to be, and this will only get worse without serious training investment...or AI.
With modern AI and efficient GPU stacks, it is now possible to generate near-real-time probabilistic simulations in a fraction of the time. By reducing a run from days to hours (or less), engineering teams are now able to massively accelerate design sprint cycles and capture that benefit through any combination of better products, faster shipping, or lower headcount.
Despite the obvious advantages of ML-defined simulation and the generalized hatred of incumbent simulation tools, adoption has been astonishingly slow. Probabilistic “surrogate” models are not mathematically rigorous or 100% deterministic. When dealing with planes, cars, and engines, even the smallest error can cost lives if not engineered correctly. “The stakes are too high,” as the argument goes, and while that is understandable, decades of that mindset have compounded into an extremely risk-averse culture where any deviation from the norm is just not worth it.
Working Thoughts
A simplstic value chain would be CAD —> Meshing —> Simulation / Solver —> Product Lifecycle Management (PLM), with all components of the stack being up for grabs. I’m particularly interested in the Meshing stage, which has been described to me by many as black magic with a uniquely strong impact on downstream solver effectiveness and overall time sink.
But what about organizations that haven’t been around for decades? Over the past three years an enormous amount if energy and captial have flowed into the macro themes of reshoring and rearmament give birth to a new generation of engineering companies. These modern startups are AI-native and unburdened by the incumbent systems of the past. It’s for this reason that I place a disprortiante weight on founders embedded in these burgeoning neo-players (e.g. SpaceX, Hadrian, Helsing) rather than the typical talent pools of Airbus and Formula 1.
The migration from equistie to attritable systmes across defense orgs is also a major driver. The cost of failure is massively lower.
There is an argument that you can’t sell to an ostich with its head in the sand. For sectors that haven’t taken the hint on modern systems you might be better off going vertical to force the industry to change / take it all for yourself. Being the best heat exchange sim is likely non-venture scale.
Relevant Companies
PhysicsX
Luminary Cloud
Nominal
Neural Concept
BeyondMath
Kernel
Revel
Flex Compute
+Incumbents like Anysys, SimScale, Cadence, Autodesk, Synopsys, MathWorks