Physics-Guided Neural Modeling Enables Reliable In-Situ Characterization of MEMS Materials

The accurate characterization of mechanical properties at the microscale is critical for the advancement and reliability of microelectromechanical systems (MEMS). Among these properties, Young’s modulus of structural materials like polysilicon plays an important role in defining the performance envelope of MEMS devices and can influence everything from resonant frequencies and sensing sensitivity to integration compatibility. However, in real-world fabrication environments, this parameter can vary significantly due to differences in deposition methods, doping concentrations, and post-processing conditions. As MEMS continue to shrink in size while expanding in application—from biomedical sensors to optical switches—the demand for precise, in-situ, and non-destructive methods of measuring material properties has grown increasingly urgent. Despite the importance of accurate property measurement, current techniques fall short in various ways. For instance, nanoindentation, while often considered the gold standard, is inherently destructive and typically limited to off-line testing and once a sample has been indented, it cannot be used in a functioning device, which fundamentally contradicts the spirit of real-time process monitoring. Optical and resonance-based methods, though non-destructive, are burdened by complex setups requiring high-precision alignment and costly instrumentation. They are also prone to errors introduced by environmental vibrations, noise, or the inability to capture true boundary conditions of embedded structures. What the MEMS community has been missing is a methodology that is simple, repeatable, compatible with wafer-level testing environments, and, crucially, does not risk damaging the device during measurement. Adding another layer of complexity is the issue of pull-in instability—a well-known failure mode in electrostatically actuated structures. Traditional electrostatic methods are appealing for their simplicity and integration into standard MEMS processes, but they often rely on voltage-induced collapse (pull-in) to derive mechanical characteristics, which inherently endangers the device. This creates a paradox: the measurement method can damage what it intends to characterize. Avoiding this while still obtaining reliable data has remained an unsolved challenge.

To this account, new research paper published in Journal of Micromechanics and Microengineering and led by Professor Zai-Fa Zhou and conducted by Dr. Zhi-peng Liang, Dr. Lin-Feng Zhao, and Dr. Qing-An Huang from the Key Laboratory of MEMS of Ministry of Education at Southeast University, researchers developed a novel in-situ method for measuring the Young’s modulus of MEMS structures using an electrostatically actuated cantilever beam that avoids pull-in instability. They integrated physics-guided neural networks trained on finite element simulation data to accurately extract mechanical properties based on the system’s voltage response. Indeed, rather than relying solely on empirical data or purely theoretical models, they crafted a hybrid approach, infusing neural networks with known physical laws to improve accuracy, reduce overfitting, and enhance interpretability.  This novel approach enables real-time, non-destructive, and reliable material characterization directly within microfabrication environments. The research team started by fabricating physical MEMS test structures using the well-established PolyMUMPs process. Each structure consisted of a cantilever beam with carefully defined geometric parameters—beam length, width, electrode spacing, and thickness—all precisely tailored to mimic the simulation conditions used in their machine learning model. These devices were not built for speculation—they were purposefully crafted to assess whether the physics-guided neural network, trained entirely on simulated data, could reliably predict material behavior in real-world conditions.

The authors’ experimental process was relatively simple: by incrementally increasing the applied voltage across the test structure and monitoring the moment of contact between the beam and electrode, the team determined the so-called contact voltage. But behind this elegant procedure was a powerful validation effort. Rather than measuring Young’s modulus directly, they input the measured contact voltage and known geometry to their trained neural network. The network, armed with its embedded physical constraints, then iteratively solved for the modulus through bisection method. This reversed the traditional paradigm—letting machine learning, not manual curve fitting, infer the material property from observed system behavior. The researchers found that both the POLY1 and POLY2 layers of polysilicon, the extracted values of Young’s modulus closely matched known benchmarks, including results from independent nanoindentation tests. The measured modulus for POLY2, for instance, came in at 165.4 GPa, which closely mirrored the 161.6 GPa obtained through nanoindentation. Perhaps more impressive was that the test structure remained fully intact throughout—no pull-in collapse, no mechanical damage. The method proved to be not only accurate, but also gentle, preserving the integrity of the MEMS component being tested. The authors also evaluated repeatability and error margins. Out of all the test cases that avoided pull-in instability, nearly 90% exhibited less than 5% error between the predicted and actual modulus values. This level of agreement underscored not just the capability of the network, but the wisdom of embedding physics into its design. It demonstrated that the model wasn’t merely memorizing patterns—it was understanding the underlying behavior.

In conclusion, the significance of the new study of Professor Zai-Fa Zhou and his colleagues  successfully developed novel hybrid approach and indeed this represents a philosophical shift in how we approach measurement at the microscale. It acknowledges that pure data-driven models often lack physical grounding, while analytical physics alone can falter in the face of complex geometries or noisy fabrication realities. With this combination, the researchers created a new method in the field of MEMS characterization that is both accurate and trustworthy. We believe one of the most important implications of this work is its potential to reshape how material properties are monitored during semiconductor manufacturing. Currently, in-situ characterization of mechanical properties is rare, largely due to destructive test methods or the impracticality of complex instrumentation. The technique presented here can be deployed directly within existing fabrication workflows, requiring only simple voltage measurements and wafer-level probing—tools already ubiquitous in cleanroom environments. That means real-time, non-destructive quality assurance for MEMS components is no longer theoretical. It’s now within reach. Moreover, the ability to predict and avoid pull-in instability through voltage margin estimation has practical safety implications. Devices that fail due to electrostatic collapse during testing are costly and slow down production and with modeling the instability threshold ahead of time, the proposed method sidesteps that risk entirely and preserve valuable test structures and streamlining development.