Flagship model launch · August 2026
From Simulating Physics to Designing the Physical Stack
Introducing Physical Design Foundation Models
Taking a breakthrough drug or an advanced battery material from a 5-milliliter lab beaker and scaling it up to a 10,000-Liter turbulent factory vessel is one of the hardest, most economically consequential problems in engineering.
Physical behavior changes dramatically with scale. A reaction that works perfectly in a test tube can easily fail, overheat, or destroy biological cells when subjected to the chaotic fluid shear of a massive industrial reactor.
Historically, the computational path between these scales has remained fragmented. Quantum chemists use one simulator; process engineers use another. Because these systems are siloed, continuous mathematical relationships break at the boundaries.
We believe a new category of foundation model is emerging to solve this: Physical Design Foundation Models.
Supported by IndiaAI
Compute that made frontier physical AI possible.
Shodh AI gratefully acknowledges the IndiaAI Mission for its GPU compute support and unwavering commitment to advancing frontier physical AI research in India.
A new model category
Three generations of foundation models
Generating information: language, code and pixels.
GenerateSimulating and optimizing individual components.
PredictDesigning coupled systems across domains and scales.
DesignGen 2 makes individual physical domains AI-native.
Gen 3 makes the coupled physical system AI-native.The category
The factory becomes computable from the desired outcome backward.
Physical Design Foundation Models are not distinguished simply by predicting physics faster, or even by performing inverse design within one domain. Their defining capability is to represent multiple interacting physical domains as one differentiable design space, so a final system objective can propagate backward across the entire stack.
If the coupled physical stack becomes differentiable, industrial scale-up stops being primarily a search problem and becomes increasingly a design problem. The factory itself becomes computable from the desired outcome backward.
Today, we are releasing our Technical Whitepaper, providing the first empirical evidence of this third generation. We have trained a foundation World Model for physical intelligence that unifies quantum thermodynamics, fluid dynamics, and biological mechanics into a single computational system.
Four controlled tests
Proving cross-scale physical intelligence
For a model to treat the entire physical stack as a single environment, controlled physical interventions must propagate correctly across scales.
Does the same neural "brain" share computation across scales?
Yes. While the model routes data dynamically based on the physics involved, we found that macroscopic factory dynamics and microscopic molecular states share 65% of their computational pathways. This suggests the model is developing a shared computational substrate across physical domains rather than routing each domain through entirely isolated pathways.
Change the chemistry. Does the factory respond correctly?
We altered a microscopic reaction-enthalpy input and tracked the response in a blinded, 250-Liter continuous reactor. The model correctly reproduced the monotonic rise in peak reactor temperature, maintaining a full-field spatial nRMSE of 0.0072 against the numerical reference.
Change the factory. Does biology respond correctly?
We generated a massive, 5,000-Liter turbulent fluid flow field at high agitation. We then mapped that macroscopic fluid shear onto a 15-micrometer biological cell membrane. The model correctly computed the mechanical force transfer, predicting an equivalent membrane stress of 312 Pa—crossing the predefined stress and area-strain rupture thresholds and producing a positive biological rupture classification.
Can the model represent competing physical mechanisms?
Increasing reactor agitation can improve oxygen transfer while simultaneously increasing damaging mechanical shear. Under a high-gas, high-kL a operating regime, the model reproduced the nonlinear trade-off: oxygen concentration had effectively plateaued between 200 and 250 RPM, while predicted near-blade shear continued to rise by 17.9%. The oxygen field passed strict quantitative validation; the shear response was directionally correct but did not pass our full spatial-correlation threshold.
The role of the neural engine
One graph. Domain structure preserved.
The point of these tests is not that Shodh has replaced quantum chemistry, CFD, or structural mechanics with one monolithic neural simulator. It has not. The important result is that heterogeneous physical representations can participate in a shared computational graph while retaining the structure required by their respective domains. Classical high-fidelity solvers remain the final numerical verifier.
The breakthrough
Cross-scale inverse design
Inverse design is already possible within individual scientific domains. A differentiable aerodynamic model can optimize a geometry for drag; a molecular model can optimize a structure for a desired property.
The harder problem is optimizing the coupled physical system.
Industrial objectives rarely depend on one solver. Factory yield depends simultaneously on molecular energetics, reaction kinetics, reactor geometry, turbulent transport, heat removal, and microscopic mechanical stress.
Because these traditionally separate physical domains are represented within one differentiable computational system at Shodh, the model can evaluate how a factory-scale objective (Jfactory) changes with respect to variables distributed across the entire physical stack.
This turns the coupled system—not merely an individual solver—into an inverse-design problem.
of proposed neural gradient directions produced an improving response.
solver-verified success across 60 feasible targets.
for a neural cross-scale proposal.
median classical solver calls bypassed.
Median solver-verified optimization time fell from more than 500 hours to approximately 5.6 hours.
From in silico to industrial execution
This is not a theoretical benchmark. We are already deploying this capability into real manufacturing environments.
When tasked with optimizing a highly exothermic, temperature-sensitive specialty chemical process, the model executed a cross-scale inverse-design query to generate a novel continuous-flow operating window. The generated parameters were cryptographically frozen, and subsequently executed by our industrial partner in a physical pilot plant.
The prospective physical execution matched the model's intent: successfully suppressing the impurity pathway from 12.3% to 3.1%, and achieving a 96.7% isolated yield in the continuous-flow pilot, compared with the process's historical 82.4% batch baseline.
Executed in an industrial pilot plant
Defining the category
From domain physics to system physics.
Shodh AI is defining Physical Design Foundation Models: models that make the coupled physical stack differentiable so engineers can optimize the system from the final outcome backward.
By restoring the broken chain between molecular discovery and factory production, we are moving AI from domain physics to system physics, and turning scale-up from a problem of empirical search into a problem of computational design.
Read the full technical whitepaper