DAC 2026 Mach42: A new generation of physics-driven AI models for analog circuit verification

Mach42 spun out of the Department of Physics at the University of Oxford in 2019 with a mission to establish software tools capable of drastically accelerating expensive physics simulations while supporting complex computational workflows. The prototype tools were successfully applied in the emulation of extreme states of matter – plasmas – and initial contracts in the fusion energy space followed.

After seed funding in 2021, the company switched focus to target a fundamental simulation bottleneck in the semiconductor market – verification of complex analog circuits. Traditional methods use non-linear equation solvers to model the behaviours of the circuits in question. These methods are highly accurate, but extremely slow, with simulation speedup largely reliant on performance improvements in the underlying hardware.

Mach42’s technology is at the forefront of a new generation of physics-driven AI models which seek to learn the underlying equations that govern complex systems. This represents a significant departure from previous generations of neural networks. Modelling the physical world is based on continuous, causal functions where the current state depends on previous states, together with impulses that occur at irregular times.

Mach42’s approach to physical systems leverages a natural, problem-adapted representation of the systems in question, leading to dramatically improved accuracy that goes far beyond existing methods. The approach relies on embedding the governing equations of the system in the network, which now has a chance of emulating the underlying physics, rather than simply capturing statistical relationships between inputs and outputs.

Mach42 accelerates the verification of analog circuits by leveraging our ability to create highly accurate surrogate models – neural networks – to augment the capabilities of traditional non-linear equation solvers. We train our surrogate models using the design testbench to create a representation of the unit-under-test.

The tools allow us to drive the target simulation environment (normally Cadence® Spectre®), and our proprietary smart sampling methods enable us to maximise the accuracy of the resulting models while minimising the training runs required. Once the model is trained, there are several exploitation routes available. It can be hosted in our environment – the Discovery Platform – where the verification engineer can examine the behaviour of the unit-under-test interactively using a GUI. In a similar vein, the model can be interrogated using scripts to conduct appropriate verification sweeps. Finally, the model can be instanced in the target simulation environment providing orders-of-magnitude simulation speedup (remember, we are executing a neural network instance, not a non-linear equation solver).

We have developed a further valuable capability that complements the simulation acceleration described above. We can generate surrogate models of a unit-under-test that produce a behavioural description of the circuit in the Verilog-A language. This enables designers to rapidly create highly accurate Verilog-A representations of their circuits automatically – a process that can take many weeks or months manually.

See more in our demo at booth #659 in the Exhibit Hall at DAC. If you’d like to book a meeting ahead of time, contact us at info@mach42.ai.

Also Read:

Beyond Transformers. Physics-Centric Machine Learning for Analog

2026 Outlook with Paul Neil of Mach42

Video EP12: How Mach42 is Changing Analog Verification with Antun Domic