The IndiaAI Mission-empanelled startup Shodh AI says its physical AI model LUCAN helped an industrial partner increase manufacturing yield from 82.4% to 96.7%, while cutting an optimisation process from more than 500 hours to 5.6 hours.
The model-generated intervention was first independently checked against classical numerical physics solvers before being taken into a physical industrial environment, according to Shodh AI’s technical whitepaper.
The company is now taking LUCAN beyond chemical and biological manufacturing, with a pilot planned with a major semiconductor equipment maker.
Unlike conventional AI models that predict the outcome of a given process, LUCAN is designed to work backwards from a desired physical outcome and identify the molecular and process variables that should change to achieve it.
Shodh AI Founder and CEO Arastu Sharma describes this as cross-scale inverse design.
From Simulation to Intervention
Industrial engineering today relies heavily on specialised tools. Molecular simulations, thermodynamics, computational fluid dynamics, biological mechanics, and process controls are typically handled separately, with engineers moving information between them and repeatedly running forward simulations.
LUCAN attempts to combine these layers.
The model uses a Sparse Mixture-of-Experts architecture that can work with different scientific representations, including molecular graphs, three-dimensional physical fields, and deforming geometries. These are mapped into a shared computational system so the model can reason across molecular, process, and equipment scales.
The distinction is between asking what will happen if a particular parameter changes and what needs to change to achieve a particular result.
As Sharma explains in a conversation with AIM, “I’m able to put the outcome that I want and ask it to tell me what parameters I should go with from the start in all of these worlds.”
The whitepaper describes the conventional engineering workflow as “propose → simulate → evaluate → adjust → repeat”. LUCAN is designed around “specify outcome → generate intervention → verify.”
That means a manufacturer could specify a target yield, for example, and have the model identify the process conditions that could move the system towards that target.
The Aarti Industries Test
Shodh AI initially focused on chemical and biological manufacturing because India already has significant industrial capacity in these areas.
One of its early industrial partners was Aarti Industries, which manufactures speciality chemicals. The company was looking to move from traditional batch manufacturing towards continuous-flow processes, where reaching the desired yield can require years of engineering and experimentation.
It used LUCAN to work backwards from the desired industrial outcome, resulting in increased yield and downsizing optimisation process.
The validation process is central to Shodh AI’s claim. The company does not position LUCAN as a replacement for established physics simulators. Instead, it uses those systems as an independent benchmark before an intervention reaches the physical world.
“These simulators exist for a reason. They are calculators. They’re amazing at their job,” Sharma asserts.
The whitepaper lays out three levels of evidence: a frozen model tested on predefined computational tasks; a model-generated intervention independently checked against high-fidelity classical physics; and a model-generated intervention physically executed in a laboratory, pilot, or industrial environment.
The Aarti result falls into the third category.
For industrial customers, raw AI benchmark performance is not the primary concern. Companies want to know whether a model agrees with the physics they already trust and, ultimately, whether its recommendations work in the real world.
That is why Shodh AI deliberately compares LUCAN against established numerical solvers rather than other AI models. The model can potentially produce a prediction much faster than a traditional simulation while also enabling inverse design and calibration with enterprise-specific data.
Why Sales Needs Physical AI
A major challenge in industrial manufacturing is that physics changes with scale.
A reaction that works in a test tube may behave differently in a larger reactor because fluid dynamics, heat transfer, mass transfer, and other physical effects change. LUCAN is designed to connect these scales rather than model them independently.
The whitepaper says information can propagate upward from microscopic energetics to reactor-scale thermodynamics, while information can also move downward from reactor-level conditions towards microscopic structural responses.
This is important for inverse design because thousands of variables can influence an industrial process.
“If an AI model is able to reason across all of these parameters at the same time—from the quantum parameters to the larger parameters—then it's not limited to a human figuring out the scaling problem,” Sharma remarks.
Shodh AI says its current model has grown from 2 billion parameters to 5 billion parameters within its IndiaAI Mission project. Sharma said the company is currently building a model in the 10–15 billion parameter range and wants to scale it towards 100 billion parameters.
The Data Security Problem
Shodh AI is also trying to address the data-security problem in deploying AI inside regulated industries.
Rather than asking customers to hand over large datasets to train the foundation model, LUCAN is trained using physics-based simulation systems and can then be deployed within an enterprise environment and calibrated with relatively small amounts of customer-specific data.
“If, say our model is 98% accurate to the physics of the world or that factory, it just needs that 2% calibration that can be done within the enterprise itself,” the founder explains.
The goal is eventually to deploy these systems inside industrial environments while keeping sensitive data within the customer's infrastructure.
From Chemicals to Semiconductors
Shodh AI's next major target is semiconductor manufacturing.
The Indian government recently connected the startup to a major semiconductor firm to conduct a pilot which makes semiconductor manufacturing equipment, but customers still need to spend significant amounts of time optimising those machines for their own manufacturing processes.
That is where Shodh AI sees an opportunity for LUCAN.
India's semiconductor ambitions make the problem particularly relevant. Building manufacturing capacity requires not only acquiring equipment but also developing the processes needed to operate it efficiently. Sharma estimates that a semiconductor prototype at the relevant scale can cost around ₹30 lakh per attempt, making physical trial and error expensive.
The company believes physical AI could reduce both the cost and time involved in developing those processes.
The Bigger IndiaAI Bet
Shodh AI was selected under the IndiaAI Mission's foundation-model programme, and Sharma says access to compute remains one of the biggest requirements for scaling LUCAN.
The company is discussing access to around 2,000 GPUs for a year.
But Sharma argues that compute is only one part of what India needs to build physical AI. He wants faster government decision-making, more willingness to take risks on deep technology, and greater access to public-sector and industrial environments where startups can test their models.
Shodh AI earlier worked with Indian Oil on a paid pilot and is exploring additional industrial deployments.
LUCAN ultimately wants to change what physical AI does on the factory floor—by not simply predicting what a physical system will do, but figuring out what humans should change to make it do something different.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.