Interview with Sushil Kumar Singh, Molex
“Agentic AI should not reduce the validation threshold”
Sushil Kumar Singh has more than 15 years of experience in fatigue, durability, vibration, fluid and thermal analysis. His work also covers high-fidelity CAE and CFD, reduced-order modelling and multidisciplinary design optimisation.
Molex LLC
E/E architectures are becoming more complex as development cycles shrink. Molex’s Sushil Kumar Singh explains how digital twins, virtual validation and agentic AI could move connector validation upstream.
As E/E architectures become more interconnected, connector development increasingly has to account for mechanical, electrical, thermal, material and manufacturing effects at the same time. Sushil Kumar Singh, Senior Structure Analyst – CAE Leader at Molex, has more than 15 years of experience in simulation-led engineering.
At the Automotive Wire Harness & EDS Conference Detroit 2026, he will join his colleague Thiruvenkata Baskaran, CAE Project Leader at Molex, for the presentation “Agentic AI Engineering: A New Paradigm for Automotive Connector Design and Validation”. Their work explores how Product Digital Twins, Virtual Validation and Agentic AI could help engineering teams identify risks earlier and coordinate increasingly complex development workflows.
Ahead of the conference, we spoke with Singh about where today’s engineering ecosystem falls short and what will be required to move more connector validation upstream.
Looking ahead five years, what will be the single biggest challenge for the wire harness and EDS industry in North America, and why?
We believe the biggest challenge will be managing rapidly increasing product and system complexity while development cycles continue to become shorter. The industry is moving towards zonal architectures, higher levels of electrification, high-speed communication, higher power density and more integrated E/E systems. At the same time, OEMs are expecting fewer design iterations and faster validation. This challenge also connects with the evolution of the work we are bringing into this conference.
Earlier discussions around the Product Digital Twin and Virtual Validation for connector development established the foundation for using trusted predictive models earlier in the product-development cycle. Our current focus is on how those capabilities can evolve further through Agentic AI. The next challenge is scaling these concepts across more functions and using them early enough that engineering risk can be identified before physical validation. Over the next five years, the differentiator will not simply be who has the most simulation tools. It will be who can build trusted, correlated, reusable and connected digital engineering capabilities that enable meaningful portions of validation to move upstream.
Where do you see the biggest gap between what next-generation E/E architectures require and what today’s development and manufacturing ecosystem can actually deliver at scale?
The biggest gap is between system-level product requirements and the fragmented way engineering data and validation capabilities are still organised today. Next-generation E/E architectures are highly interconnected. Mechanical behaviour affects electrical performance; thermal conditions affect materials; manufacturing variation affects sealing and terminal performance; and vibration can affect PCB, connector and interface reliability. However, the engineering evidence behind these product features is still developed in silos, limiting traceability and cross-functional decision-making. This is where we see the Product Digital Twin as an important foundation.
What makes an agentic AI engineering workflow fundamentally different from conventional optimisation, reduced-order modelling or an AI assistant used within CAE?
We see Agentic AI as the next evolution of the Digital Twin and Virtual Validation journey. Traditional optimisation searches within a design space defined by an engineer. Reduced-order models mainly accelerate computationally expensive simulations. AI assistants can help engineers write scripts, generate models or interpret results. Agentic AI potentially goes further by operating across the engineering workflow itself. An engineering agent could potentially help to understand requirements, select the appropriate Digital Twin capability, execute simulations, assess requirements, identify risks, explore alternatives, rerun analyses, document evidence and recommend the next engineering action.
That is fundamentally different because the AI is no longer supporting only one task. It is coordinating multiple engineering activities towards an objective. However, my view is that Agentic AI should sit on top of trusted predictive capabilities, not replace them. The Digital Twin provides the physics and product knowledge. Virtual Validation provides the evidence and acceptance framework. Agentic AI provides the orchestration and autonomous decision workflow.
How do you prevent an AI agent from finding an apparently optimal design that satisfies its target metric but violates physics, manufacturability, material limits or long-term reliability elsewhere?
This is where the Digital Twin framework becomes essential. A connector is a multifunctional system. Optimising only one response can easily create a problem somewhere else. For example, reducing connector mate force may appear beneficial, but the same design change could affect terminal normal force, electrical contact reliability, retention, latch robustness, sealing compression, manufacturability, dimensional robustness or long-term relaxation. An Agentic AI workflow therefore cannot be built around a single optimisation target. It needs to operate within a multifunctional engineering constraint framework.
This is where the Product Digital Twin becomes essential. The Digital Twin should provide the validated physics models, material behaviour, manufacturing variation and product-level relationships, while the Virtual Validation framework defines the acceptable performance envelope and evidence required before a design is considered valid. One of the important gaps we still need to overcome is making AI models more physics-aware, so they do not simply learn statistical relationships from data and extrapolate into physically unrealistic regions. This is where approaches such as physics-informed neural networks, or PINNs, are relevant.
By embedding governing equations, boundary conditions, conservation laws or other physics constraints directly into the learning process, PINNs can help constrain AI predictions towards physically consistent behaviour. However, PINNs alone are not sufficient. Manufacturability, material limits, dimensional variation, contact behaviour, durability and long-term reliability still need to be represented through validated engineering models and explicit constraints.
What data foundation is required to combine high-fidelity CAE, material characterisation, physical testing, environmental ageing and manufacturing variation, particularly when training data are sparse or outside the conditions previously observed?
A credible Product Digital Twin requires much more than a repository of simulation results. It needs a traceable engineering data foundation. For connector systems, performance can depend on factors such as material, fibre orientation, weld lines, conditioning, temperature, ageing, friction, tolerance and interface behaviour. The data therefore needs to preserve context such as geometry revision, material grade and condition, manufacturing and process inputs, environmental condition, test configuration, boundary conditions, simulation assumptions, correlation evidence and uncertainty.
This directly extends the Virtual Validation framework Thiru presented at Bordnetzkongress, where material characterisation, process effects, physics-based simulation and robustness were treated as interconnected pieces rather than independent activities. AI should complement the physics. One of the most important capabilities of an engineering AI system should be recognising when it is operating outside its validated domain. In engineering, being able to say, “there is not yet enough evidence to make this prediction reliably”, can be just as important as producing a prediction.
What level of correlation, uncertainty quantification and traceability would you require before an AI-driven workflow could safely eliminate even one physical prototype or validation iteration from connector development?
We do not believe one universal correlation percentage should define whether physical testing can be replaced. The required evidence should depend on the functional requirement, model maturity and consequence of an incorrect prediction. Before reaching Virtual Validation, I would expect three things. First, correlation: the model should reproduce the relevant physical behaviour, not just a single peak number. Depending on the application, that could mean force-displacement behaviour, failure mode, strain distribution, modal response, leakage behaviour, temperature response or durability trend.
Second, uncertainty and robustness: the model should account for relevant variation such as material properties, manufacturing tolerances, fibre orientation, environmental conditioning, process variation and test variability. Third, traceability: as engineering workflows become more autonomous, we need even stronger visibility into how decisions were made, what data and models were used, and whether the required validation evidence was satisfied. Agentic AI should not reduce the validation threshold; it should accelerate the generation and traceability of the evidence required to meet it.
Finally, what do you personally hope to take away from the Automotive Wire Harness & EDS Conference Detroit 2026?
Personally, I hope to make the EDS Conference a platform to share knowledge, learn from industry peers and better understand customer needs, expectations and emerging market trends.
