DeepControl Funding: The 'Explainable' Route for Physical AI Finally Has Industry Proof [Analysis]

Industrial AI has been hyped for years, but most projects get stuck after the POC—not because accuracy is insufficient, but because "black boxes are untrustworthy." Workshop supervisors would rather use PID loops than let a model outputting probability distributions directly control valves. Seeing this funding report on DeepControl, I'm particularly interested in how they use physical mechanisms to solve this trust issue.

Core info: Based on their proprietary PhyAI engine, deeply coupling physical mechanisms with AI, generalization error is below 3%, and they've achieved L4/L5 level autonomous closed-loop control. This isn't running demos in simulation environments; it's actually integrated into real systems of 360 top clients like TSMC, Tencent, and ByteDance, covering over 300,000 devices. Achieving commercial profitability indicates this tech route genuinely works in specific scenarios like electromechanical energy systems.

I agree with the "physics-based foundation" approach. From first principles, the response of systems like building HVAC and data center temperature control is essentially governed by thermodynamics and fluid dynamics equations. Hard-fitting these with data-driven methods is inefficient and non-transferable. By embedding mechanisms as prior knowledge into the network, DeepControl adds "physical conservation law" constraints to the AI, making decisions naturally interpretable and generalizable. This is similar to the Nature paper using PINNs for flow field reconstruction, except they're doing real-time control.

However, a few points are debatable. First, L4/L5 levels lack clear definitions in industry; the report mentions a "generational leap from L2/L3 to L4/L5"—are there redundant safety circuits in actual deployment? Second, moving from device-level control to "source-grid-load-storage integrated global energy scheduling" increases complexity exponentially. With limited physical model coverage, can they maintain 3% error then? Third, overseas markets—AC load characteristics in Southeast Asia and the Middle East differ greatly from domestic ones. Does migrating physical models require recalibrating parameters?

Finally, a question: In strongly coupled, multi-timescale scenarios like grid-level scheduling, what exactly is the incremental engineering advantage of Physical AI compared to traditional Model Predictive Control (MPC)?

https://36kr.com/p/3887726503688968?f=rss