Sumanth Murthy is an AI research engineer with 8+ years
building machine-learning systems from first principles to
production. He is currently Evaluation Lead for AI Systems at
Lenovo and CTO of OptivAI, where he designed physicsinformed world models and agentic AI platforms deployed
across industrial facilities. Previously he led computer vision
and research for a multimodal biological foundation model at
CYTED, work published in Nature Communications (in
collaboration with Microsoft Research) and deployed across
50+ NHS organisations. His interests span physics-informed
learning, foundation models, reinforcement learning, and
rigorous AI evaluation. He holds an MSc in Robotics from the
University of Toronto Institute for Aerospace Studies, serves
as a peer reviewer for ICLR, and is a UK Aspen Society
Fellow.

When More Physics Hurts: The Non-Monotonic Value of Governing Equations in Sim-to-Real Reinforcement Learning

Reinforcement-learning agents trained in simulation routinely
fail on deployment because purely neural world models learn
simulator artefacts rather than the underlying physics. We
present a physics-informed world model that addresses this by
encoding known governing equations as a structural prior
while a neural network learns only the residual deviation from
theory — the model cannot unlearn the physics because it
never learns it. Built on a DreamerV3-style recurrent statespace model for industrial thermodynamic control, the
approach reduced the sim-to-real gap from 30–50% (pure
neural) to 5–6% on identical data.
Our central and, to our knowledge, underreported finding is
that the value of the physics prior is non-monotonic in its
strength. Across 40+ systematic ablations we characterise a
prior-strength curve with a clear interior optimum: too weak,
and the model overfits simulation artefacts; too strong, and it
cannot fit genuine real-world deviations such as transient
dynamics, occupancy effects, and equipment degradation.
This mechanism is visible only from the shape of the curve,
not from any single experiment — a caution for the common
assumption that more physics is always better.
We further show that ensemble-disagreement uncertainty,
calibrated against physical ground truth, enables safe
uncertainty-gated abstention during staged deployment. The
resulting system ran in production across five facilities at
99.9% reliability over one year, delivering an 18.9%
efficiency improvement. We discuss design principles for
deciding which parts of a physical system to encode
structurally versus learn empirically.