1) Quantum Optics & Photonics, 2) Wide Band Gap (WBG) Devices & Design Automation,
3) MEMS With Physics-In-The-Loop Inverse Design, 4) AI, Control & Energy Systems, and 5) Photovoltaic Modeling
Title: Overcoming the 3 dB squeezing extraction limit in silicon carbide microcombs with a photonic molecule
We developed an open and reproducible design pipeline for engineering strongly squeezed quantum states in soliton-crystal microcombs based on 4H-silicon-carbide-on-insulator resonators. The framework integrates full-vector finite-element modeling, Lugiato–Lefever equation simulations, and linearized Heisenberg–Langevin analysis to establish a complete workflow from material parameters to experimentally detectable quantum noise. To overcome the inherent 3 dB squeezing extraction limit of conventional resonator coupling, we introduced an auxiliary Purcell-extraction ring that selectively enhances access to squeezed below-threshold modes. The optimized design achieves 8.5 dB of detectable squeezing over a 1.74 GHz bandwidth at a fixed pump power, while generating dominant squeezed supermodes and entangled quadrature lattices. This work provides a scalable and openly available computational framework for designing integrated quantum photonic sources and advancing continuous-variable quantum information technologies.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: Imaging Single Vortices and Their Losses in Tantalum Superconducting Circuits With a Pick-and-Place Diamond Quantum Sensor: An End-to-End Simulation Study
We developed an open-source simulation framework for a nanoscale quantum sensing platform to enable in situ characterization of vortex dynamics in superconducting tantalum circuits. By combining time-dependent Ginzburg–Landau simulations, physics-informed inversion, and nitrogen-vacancy (NV) center magnetometry, the framework accurately localizes vortices, identifies pinning states, and quantifies dissipation, providing a powerful tool for studying vortex-induced losses in superconducting quantum devices.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: SPARQ—spiking physics-in-the-loop autonomous reinforcement triage of quantum emitters: a closed-loop, event-driven machine-learning framework for characterizing solid-state single-photon sources
We propose SPARQ (Spiking Physics-in-the-Loop Autonomous Reinforcement Triage of Quantum Emitters), a machine-learning framework that transforms Hanbury Brown–Twiss (HBT) autocorrelation measurements into a closed-loop, event-driven, and transferable instrument for autonomous quantum-emitter characterization. At its core is a stochastic differentiable digital twin of the HBT measurement chain, validated against numerically exact master-equation solutions. A spiking neural network trained through the twin achieves the accuracy of computationally intensive multi-start fitting while reducing measurement latency by 6× and enabling energy-efficient, anytime inference. Adjoint optimization jointly tunes excitation power, correlation window, and estimator, improving performance under photon-starved conditions. A soft actor–critic agent with prioritized experience replay autonomously prioritizes emitter measurements, accelerating large-area screening by 1.7×. The framework transfers zero-shot to publicly available quantum-dot data and, through platform-randomized graph conditioning, generalizes across diverse emitter platforms, with benchmark datasets and trained models released openly.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: PILOT-Q: physics-in-the-loop training and photon-budget-aware closed-loop operation of delocalized photonic neural inference at the standard quantum limit
We developed PILOT-Q (Physics-in-the-Loop Training and Confidence-Gated Operation at the Quantum Limit), a framework for robust and photon-efficient delocalized photonic inference under single-photon operating conditions. At its core is a differentiable stochastic digital twin of the optical broadcast chain that accurately models standard quantum limit (SQL) shot noise, dark counts, residual weight-trim error, and quantization. Training through this twin substantially improves photon-starved inference accuracy—from 51.6% to 78.6% at one photon per MAC—while reducing the photon budget required to achieve matched accuracy by 1.4–3.1× across three public benchmarks. A confidence-gated controller adaptively reallocates photons to uncertain inputs, although oracle analysis shows that physics-aware training, rather than run-time adaptivity, provides the dominant performance gains. The framework establishes a practical approach for energy-efficient photonic AI at the quantum limit, with the digital twin, trained models, and experimental measurements released openly.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: FabGAN-ID: Learned Generative Process Twins in a Fully Differentiable Fabrication Loop for Yield-Aware Inverse Design of Multilayer Photonic Sensor Front-Ends
We propose FabGAN-ID, a fabrication-aware inverse-design framework that replaces conventional Gaussian process models with a learned generative fabrication twin inside a fully differentiable optimization loop. A conditional, moment-matched Wasserstein GAN learns the joint distribution of fabrication errors from only 400 historical process traces, accurately capturing systematic, correlated, and heavy-tailed variations that dominate manufacturing yield. Integrated with an exact differentiable physics solver, the learned twin enables direct optimization of conditional value-at-risk (CVaR), improving the yield floor of a 532-nm fluorescence-rejection notch filter by 7.2% over the nominal design and outperforming all Gaussian-based robustification methods on the true fabrication process. The differentiable framework also supports analytic policy-gradient optimization for specification-conditioned correction policies, where model-free reinforcement learning fails, while demonstrating superior generalization and efficiency compared with neural surrogate models. The complete benchmark dataset, comprising 48,300 spectra and 400 fabrication traces, together with the trained models, is released openly.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: Physics-Informed Machine Learning and Adjoint Co-Design of a Monolithic E-Mode GaN-on-SOI Fractional-N PLL With HEMT Varactors for Extreme-Temperature Timing
We developed a physics-informed machine learning framework for the automated co-design of a monolithic fractional-N phase-locked loop (PLL) in 200-V GaN-on-SOI technology. By integrating physics-informed neural networks, adjoint-based gradient optimization, and graph-neural reinforcement learning, the framework optimizes circuit performance, thermal robustness, and manufacturability. The resulting PLL demonstrates substantially reduced phase error and jitter, improved lock yield, compact layout implementation, and reliable fractional-N operation across a wide temperature range.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: Physics-Guided Graph Reinforcement Learning with an Adaptive Fuzzy Reward for Transferable Analog Transistor Sizing
In collaboration with GlobalFoundries Inc., we developed a graph-based reinforcement learning framework for automatic transistor sizing that combines graph convolutional networks, interval type-2 fuzzy reward learning, and physics-guided adjoint optimization. Unlike conventional approaches that rely on black-box circuit simulation and fixed weighted objectives, the proposed framework incorporates uncertainty-aware reward modeling and physics-informed gradients to improve optimization quality and sample efficiency. Validated across multiple technology nodes and amplifier benchmarks, it consistently outperforms existing Bayesian optimization and reinforcement learning methods while demonstrating strong transferability across circuit topologies.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: PARL-ID: A Fabrication-Aware Physics-Informed Adjoint Framework With Reinforcement Learning in the Loop for Inverse Design of CMUT and Photonic Sensors
We developed PARL-ID, a unified physics-informed inverse design framework that integrates PINNs, adjoint optimization, and reinforcement learning to automate robust sensor design. Validated on CMUTs and photonic devices, it improves optimization efficiency while producing fabrication-tolerant designs that outperform conventional data-driven inverse design methods.
Paper: Under review
Code: GitHub-Tanvir-Mahmud-Mahim
Title: Hierarchical Inverse Design Framework for Unit-Cell CMUTs With Attentive Gated Recurrent And Fully Connected Dense Layers
Inverse design of CMUTs brings a new dimension to automating the microelectromechanical device profile optimization.
In a published paper of mine in IEEE Sensors Journal, we developed a hierarchical machine learning (ML)-based scheme where a probabilistic search algorithm is derived to find the optimum ML architectural hierarchy, which is then used to learn the objective function of a unit cell CMUTs (i.e., maximizing the membrane displacement profile).
I have provided snapshots of the optimized device profile of the unit-cell CMUTs and the developed hierarchical ML scheme (taken from my published paper in the IEEE Sensors Journal).
Paper: IEEE-Sensors-Journal-Article
Title: Adaptive Takagi-Sugeno-Kang Fuzzy Inference with Reinforcement Learning to Withstand Extreme Low Voltages of DFIG-WT
Ensuring grid stability under fault conditions is a critical challenge in modern renewable energy systems. In a published paper in the Engineering Science and Technology, an International Journal (JESTCH), we developed an adaptive intelligent control scheme for DFIG wind turbines to withstand extreme low voltage events caused by high wind penetration and electromechanical transients.
The core of the proposed framework couples a Takagi–Sugeno–Kang Fuzzy (TSKF) conditional inference engine with Double Q-learning reinforcement learning. The TSKF module infers the phase angle and modulation indices of a STATCOM connected to the DFIG, leveraging its interpretability and nonlinear mapping capability.
I have provided snapshots of the control flow diagram and DFIG transients (taken from my published paper in the JESTCH).
Paper: Elsevier-JESTCH-Article
Title: Adaptive Fuzzy Attention Inference to Control A Microgrid Under Extreme Fault On Grid Bus
Reinforcement learning (RL) is a critical player in adaptive/dynamic systems. Within the plethora of RL, a special kind of RL is the Q-learning technique, where the actions and states are mapped elegantly, with corresponding rewards. The power of RL is immense in diverse fields.
In a published paper of mine in IEEE Transactions on Fuzzy Systems (IEEE TFS), a double Q learning scheme with customized prioritized reward with the help of an experience buffer reply mechanism was implemented. This modification allowed for more robust control of a complex microgrid system with a fuzzy attention inference framework. A snapshot of the results relating to the adaptability is attached to this post (taken from my paper in IEEE TFS).
Paper: IEEE-TFS-Journal-Article
Title: Weather Responsive Multidimensional Photovoltaic Efficiency Model for Simulation of Custom-Built Bifacial Panel
A paper of mine in the IEEE Journal of Photovoltaics, where part of the contribution was to integrate additive weather parameters into the photovoltaic efficiency rating model.
I was partly motivated by the beautiful derivation from the one-diode model towards the photovoltaic efficiency model derived by Anton Driesse et al. (paper: A New Photovoltaic Module Efficiency Model for Energy Prediction and Rating, link: Anton-IEEE-JPV), where their ADR model uses irradiance (g) and ambient temperature (T).
In my paper, additional parameters (i.e., air pressure and air humidity) are carefully crafted to provide a more robust modeling of the photovoltaic efficiency rating scheme. Besides, the model was validated for different PV cell technology-based modules; i.e., (a) CdTe-S4V3 (FirstSolar), (b) n-type mono c-Si-LG320N1K-A5 (LG), and (c) HIT c-Si- VBHN325SA16 (Panasonic). I have attached snapshots of my proposed model's equations and validation across different PV cell technology-based modules efficiency; from my paper that was published in the IEEE Journal of Photovoltaics (IEEE JPV).
Paper: IEEE-JPV-Journal-Article
Title: Review of Mono- and Bifacial Photovoltaic Technologies: A Comparative Study
Next-generation bifacial PV cells are vital in developing high-efficiency modules. In a paper of mine in the IEEE Journal of Photovoltaics (IEEE JPV), the likes of technologies such as TOPcon, Si-heterojunction, and so on are presented in detail.
Paper: IEEE-JPV-Journal-Article
Title: Review of the Challenges and Prospects in Agrivoltaics
Dual use of land is essential, which can be achieved by agrivoltaics that provide economic benefits to end users. This articles reviews the recent prospects of agri-photovoltaics (agri-PV) across continents. The existing standards and pilot projects to standardize new codes and regulations are discussed. The state-of-the-art agri-PV scheme's benefits and lackings are presented. Community acceptance and policy barriers are also discussed where prospective solutions presented in the literature are organized.
Paper: WILEY-AESR-Journal-Article
