Tanvir M. Mahim

Research

Device physics modeled from first principles, and inverted to determine the designs that meet a specification.

My research develops the device physics of microcombs and quantum sensors, of monolayer transistors and microelectromechanical transducers, and of GaN circuits and analog blocks. Each device is modeled from first principles through to the quantities that an experimental measurement or a circuit simulation reports. The models are then inverted: given a target specification, the objective is to determine the geometry that achieves it and to establish whether that geometry remains viable under realistic process variation. Because this inversion requires the governing physics to be differentiable, the simulators are constructed to admit gradients throughout, and each is paired with an optimizer, whether adjoint, generative or reinforcement-learned, that searches the design space directly under the constraints imposed by a realistic fabrication process.

Work marked Under review is currently in peer review. The Publications page lists only accepted records. Code and datasets are released openly as each project reaches maturity.

Quantum optics and photonics

Conducted with Dr. A. S. M. Mohsin, Department of EEE, BRAC University, since July 2025. This work develops differentiable models of quantum photonic hardware, including squeezed-light sources, color-center magnetometers and single-photon instrumentation, together with learned controllers that close the loop around them.

Exceeding the 3 dB squeezing-extraction limit in silicon-carbide microcombs

4H-SiC-on-insulator · soliton crystals · continuous-variable quantum optics|Under review

Photonic-molecule geometry with a main soliton-crystal ring coupled to an auxiliary Purcell-extraction ring, the comb-tooth and squeezed-vacuum mode structure, and finite-element mode profile and dispersion engineering of the 4H-SiC waveguide core.
(a) Photonic molecule: a two-FSR soliton-crystal main ring side-coupled to an auxiliary extraction ring. (b) Even comb teeth, and the odd squeezed-vacuum modes the auxiliary ring reaches. (c–e) Finite-element mode profile of the 1.85 µm × 500 nm 4H-SiC core, and the dispersion engineering that fixes the operating geometry.

This work develops an open and reproducible pipeline for engineering strongly squeezed quantum states in soliton-crystal microcombs on 4H-silicon-carbide-on-insulator. Full-vector finite-element modeling, Lugiato–Lefever simulation and linearized Heisenberg–Langevin analysis are chained into a single workflow running from material parameters through to experimentally detectable quantum noise. Conventional resonator coupling limits extractable squeezing to 3 dB. Introducing an auxiliary Purcell-extraction ring, which selectively opens a channel to the below-threshold squeezed modes, raises this limit. The optimized design reaches 8.5 dB of detectable squeezing across a 1.74 GHz band at fixed pump power, while generating dominant squeezed supermodes and entangled quadrature lattices.

Imaging single vortices in tantalum superconducting circuits

Diamond NV magnetometry · time-dependent Ginzburg–Landau · superconducting-qubit loss|Under review

Three-panel figure: simulated vortex matter in a patterned tantalum film, a pick-and-place diamond micromembrane carrying shallow nitrogen-vacancy centers above the film, and the reconstructed nitrogen-vacancy plane field map.
Target: vortex matter in a patterned Ta film, from time-dependent Ginzburg–Landau simulation at measured parameters. Platform: a pick-and-place diamond micromembrane with NV centers 6 nm deep, held at 25 nm standoff. Outcome: an NV-plane field map inverted to 13 nm vortex localization, with T₁ and T₂ mapping onto vortex drag and pinning.

This work develops an open-source simulation framework for a nanoscale quantum sensing platform that characterizes vortex dynamics in superconducting tantalum circuits in situ. Time-dependent Ginzburg–Landau simulation, physics-informed inversion and nitrogen-vacancy magnetometry combine to localize individual vortices, identify their pinning states and quantify the dissipation they produce. This provides a route to the systematic study of the vortex-induced losses that limit superconducting quantum devices.

SPARQ: autonomous triage of solid-state single-photon emitters

With the University of Memphis, USA · spiking networks · Hanbury Brown–Twiss instrumentation|Under review

SPARQ architecture: a confocal emitter field feeding Hanbury Brown-Twiss detection, a spiking front-end and a soft actor-critic agent; the physics-in-the-loop training loop with a stochastic twin and WGAN-GP critic; and level-structure graph conditioning across NV, hBN, GaN and SiV platforms.
(a) The closed loop: confocal emitter field → HBT detection → event-driven spiking front-end → soft actor–critic agent with prioritized replay, actuating a per-site photon budget. (b) Physics-in-the-loop training against a stochastic twin, with a WGAN-GP critic as sim-to-real diagnostic. (c) Level-structure graph conditioning that carries the estimator zero-shot across NV, hBN, GaN and SiV platforms.

SPARQ reformulates Hanbury Brown–Twiss autocorrelation measurement as a closed-loop, event-driven instrument rather than an offline fitting procedure. At its center is a stochastic differentiable digital twin of the HBT measurement chain, validated against numerically exact master-equation solutions. A spiking neural network trained through that twin matches the accuracy of computationally intensive multi-start fitting while cutting measurement latency by 6× and supporting energy-efficient anytime inference. Adjoint optimization tunes excitation power, correlation window and estimator together, an effect that is most pronounced under photon-starved conditions. A soft actor–critic agent with prioritized experience replay then chooses which emitter to measure next, accelerating large-area screening by 1.7×. The framework transfers zero-shot to public quantum-dot data and, through platform-randomized graph conditioning, generalizes across emitter platforms it never saw in training.

PILOT-Q: photon-efficient neural inference at the standard quantum limit

With the University of Memphis, USA · photonic computing · shot-noise-limited operation|Under review

PILOT-Q framework: delocalized photonic inference from a central light server to drone, camera and IoT nodes; one in-physics layer annotated with trim error, shot noise, dark counts and ADC quantization; photon-budget-aware training; and closed-loop confidence-gated operation.
(a) Delocalized photonic inference: one central light server broadcasting weight-encoded light to remote nodes. (b) A single in-physics layer with its four impairments: weight-trim error, standard-quantum-limit shot noise, dark counts and b-bit ADC quantization. (c) Photon-budget-aware training through the stochastic twin. (d) Confidence-gated operation, re-exposing only uncertain inputs at higher photon budget.

PILOT-Q addresses inference on optical hardware in which each multiply–accumulate operation is carried by only a few photons. A differentiable stochastic twin of the optical broadcast chain models standard-quantum-limit shot noise, dark counts, residual weight-trim error and quantization. Training through it lifts photon-starved accuracy from 51.6% to 78.6% at one photon per MAC, and reduces the photon budget needed for matched accuracy by 1.4–3.1× across three public benchmarks. A confidence-gated controller reallocates photons to uncertain inputs. Oracle analysis shows that the dominant gain arises from physics-aware training rather than from run-time adaptivity. This is a negative result, and it is reported explicitly because it directs subsequent effort toward the training procedure rather than the run-time controller.

FabGAN-ID: learned fabrication twins for yield-aware photonic design

Generative process models · differentiable CVaR optimization · photonic sensor front-ends|Under review

FabGAN-ID pipeline from process traces through a GAN fabrication twin and a differentiable CVaR loop to a yield-qualified sensor, with plots showing the learned twin reproducing heavy error tails and the yield tail lifted by 7.2 percent.
Process traces feed a conditional GAN fabrication twin, which sits inside a differentiable conditional-value-at-risk loop. Left: the learned twin reproduces the heavy tails of the true thickness-error distribution that Gaussian models miss. Right: the 5% CVaR yield floor of a 532 nm notch filter, lifted by 7.2 points over the nominal design.

Robust photonic design conventionally assumes Gaussian process variation. FabGAN-ID replaces that assumption with a learned generative fabrication twin placed 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, capturing the systematic, correlated and heavy-tailed variation that governs yield in practice. Coupled to an exact differentiable physics solver, it enables direct optimization of conditional value-at-risk, improving the yield floor of a 532 nm fluorescence-rejection notch filter by 7.2% over the nominal design and outperforming every Gaussian-based robustification method on the true fabrication process. It also supports analytic policy-gradient optimization for specification-conditioned correction policies, where model-free reinforcement learning fails. Released with 48,300 spectra and 400 fabrication traces.

2D quantum materials and MEMS

Conducted with Prof. Md. Mosaddequr Rahman, Department of EEE, BRAC University, since July 2024. This work treats monolayer transition-metal-dichalcogenide devices from first principles through to circuit-level metrics, and applies physics-in-the-loop inverse design to micromachined transducers.

Two Raman phonons that measure edge charge in monolayer nanoribbon transistors

Monolayer TMDs · frozen-phonon DFT · tip-enhanced Raman · width scaling|Under review

Tip-enhanced Raman measurement of a monolayer nanoribbon transistor, spectra comparing ribbon center and edge, a plot separating edge charge from edge strain, and normalized on-current versus nanoribbon width for two gate stacks.
(a) Tip-enhanced Raman on a 1H-monolayer nanoribbon channel, with the edge damage halo and band-edge profile inset. (b) At the ribbon edge the A′₁ mode shifts by 0.5 cm⁻¹ while 2LA(M) does not. That is the signature that separates charge from strain. (c) Edge: charge, no strain. Interior: strain, no charge. (d) Width scaling: critical width falls from 252 nm on a 90 nm SiO₂ gate to 18 nm on a thin high-κ gate.

Edge charge and edge strain both shift the Raman lines of monolayer transition-metal dichalcogenide nanoribbons, and distinguishing between them is the step that converts a spectroscopic observation into a quantitative electrostatic diagnostic. First-principles frozen-phonon calculations show that the 2LA(M) mode is strongly strain-sensitive while A′₁ responds mainly to carrier density. Two phonons give two independent readouts. Applied to tip-enhanced Raman data, the method quantifies edge charge accumulation and finds negligible edge strain, which explains the observed width-scaling behavior. Edge charge governs the depletion-to-enhancement transition on thick-oxide devices, while on high-κ gated transistors the critical width is set by patterning-induced damage rather than intrinsic defects. This distinction determines the appropriate corrective action in a process flow.

Strain and ferroelectricity together in WSe2 nonvolatile logic

van der Waals ferroelectrics · two-valley Boltzmann transport · nonvolatile circuits|Under review

Strained p-type MFMIS ferroelectric field-effect transistor stack with monolayer WSe2 channel and CuInP2S6 gate, a valence-band diagram showing suppressed intervalley scattering under compression, and the multiscale simulation chain.
(a) Strained p-type MFMIS stack: monolayer WSe2 channel, h-BN, floating gate and a CuInP2S6 van der Waals ferroelectric under biaxial compression. (b) Compression lowers the Γ valley and suppresses intervalley scattering. (c) The multiscale chain, from strained two-valley transport through multidomain ferroelectric kinetics to nonvolatile-logic metrics.

This work develops a multiscale simulation framework for a strain-engineered ferroelectric field-effect transistor built from a monolayer WSe2 channel and a CuInP2S6 van der Waals ferroelectric gate stack, coupling a calibrated two-valley Boltzmann transport model to self-consistent electrostatics and multidomain ferroelectric dynamics. One percent biaxial compressive strain more than doubles hole mobility and raises retained on-current by 1.8× while leaving the memory window intact. This indicates that strain and ferroelectric polarization act as independent design parameters rather than competing ones. At circuit level, the resulting complementary nonvolatile inverters and latches recover their state after complete power loss, cut worst-case static power by eight orders of magnitude against prior CIPS latches, and improve the energy–delay product by 2.3×.

PARL-ID: fabrication-aware inverse design across MEMS and photonics

Physics-informed neural networks · neural adjoint · CVaR reinforcement learning|Under review

PARL-ID three-stage architecture: a multi-physics physics-informed neural network forward surrogate, a neural-adjoint inverse engine, and a reinforcement-learning fabrication loop, validated on a unit-cell CMUT testbench and photonic benchmarks.
Three stages: a multi-physics PINN forward surrogate with hard boundary-condition encoding, a neural-adjoint inverse engine performing projected gradient descent over the fabrication-feasible set, and a GCN-SAC fabrication loop whose reward is the tail risk (CVaR) over sampled process corruptions. One architecture, two sensor domains: MEMS ultrasonics and integrated photonics.

PARL-ID unifies physics-informed neural networks, adjoint optimization and reinforcement learning into a single inverse-design framework for robust sensor design. Validated on CMUTs and on photonic benchmarks, it improves optimization efficiency while producing fabrication-tolerant designs that outperform conventional data-driven inverse design. It is the direct successor to the published CMUT framework described below: whereas that work identified the optimal network hierarchy for a nominal design, the present framework optimizes for designs that remain within specification under process variation.

Hierarchical inverse design of unit-cell CMUTs

Attentive gated recurrent networks · membrane-displacement maximization|Published · IEEE Sensors Journal 2025

Hierarchical inverse-design network: unit-cell CMUT thickness parameters enter a stack of gated recurrent layers, pass through an attention dot-product block and fully connected dense layers, and emerge as an optimized device profile.
The hierarchical inverse-design network. Unit-cell CMUT thickness parameters enter a stack of GRU layers, pass through an attention block, and emerge from fully connected dense layers as an optimized device profile. Figure from the IEEE Sensors Journal paper.

Inverse design automates a procedure that is otherwise a manual parameter sweep over microelectromechanical device profiles. In this published work a probabilistic search algorithm derives the best machine-learning architecture, which turns out to be attentive gated recurrent layers feeding fully connected dense layers. That network maps a target acoustic response back to unit-cell CMUT geometry and maximizes membrane displacement without the exhaustive finite-element sweeps that unit-cell design normally demands.

Wide-bandgap devices and design automation

Conducted with Dr. Nadim Chowdhury, Department of EEE, BUET, since May 2025. This work applies gradient-based design to GaN circuits and develops reinforcement-learning methods that transfer across process nodes.

Co-designing a monolithic GaN-on-SOI fractional-N PLL

200 V GaN-on-SOI · E-mode HEMT varactors · extreme-temperature timing · design partnership with the Palacios Group, MIT|Under review

Three panels: manual GaN PLL design failing to reach its 50 MHz target, the co-design engine combining a PINN varactor surrogate with adjoint gradients and reinforcement learning, and the resulting fractional-N synthesizer locking across 218 to 423 kelvin.
Problem: manual design, with measured parasitics, mismatch and leakage, misses the 50 MHz target and only unlocks above 233 K. Engine: a PINN varactor surrogate C(V,T), a differentiable fractional-N loop and noise model, adjoint gradients over twelve design parameters, a GCN-SAC certificate and WGAN-GP variability, all on an open toolchain of OpenVAF, ngspice and gdstk. Result: a synthesizer locking 10/10 at 46.5 MHz across 218–423 K.

This work develops a physics-informed machine-learning framework for the automated co-design of a monolithic fractional-N phase-locked loop in 200 V GaN-on-SOI technology. Physics-informed neural networks, adjoint-based gradient optimization and graph-neural reinforcement learning jointly optimize circuit performance, thermal robustness and manufacturability. These three objectives are conventionally traded off manually and in sequence. The resulting PLL shows substantially reduced phase error and jitter, improved lock yield, a compact layout implementation, and reliable fractional-N operation across a wide temperature range.

Transferable analog transistor sizing with graph reinforcement learning

With GlobalFoundries, Inc., Santa Clara, USA · interval type-2 fuzzy rewards · 180/130/65/45 nm|Under review

Transistor sizing loop: a circuit graph encoded by a graph convolutional network feeds a soft actor-critic agent driving ngspice across open PDKs, with an interval type-2 fuzzy reward and a physics-in-the-loop adjoint supplying exact gradients to the actor.
Devices are nodes and nets are edges. A GCN encoder feeds a soft actor–critic agent that drives ngspice across open PDKs at 180/130/65/45 nm. An interval type-2 TSK fuzzy reward with a non-zero footprint of uncertainty shapes the return, while a physics-in-the-loop adjoint supplies exact gradients of a differentiable figure of merit straight to the actor. The pretrained encoder transfers to new topologies.

Developed with GlobalFoundries, Inc., this work presents 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. Whereas conventional approaches rely on black-box circuit simulation and fixed weighted objectives, the present framework incorporates uncertainty-aware reward modeling and physics-informed gradients, improving both optimization quality and sample efficiency. Validated across multiple technology nodes and amplifier benchmarks, it consistently outperforms existing Bayesian-optimization and reinforcement-learning methods while transferring across circuit topologies.

Machine learning for control and energy systems

Conducted with Dr. A. H. M. A. Rahim (retired December 2024) between 2023 and 2024. This work develops controllers that adapt their own inference rules, applied to grid-connected machines under fault conditions.

Fuzzy inference with reinforcement learning for DFIG low-voltage ride-through

Takagi–Sugeno–Kang inference · doubly-fed induction generators · extreme grid sags|Published · JESTECH 2026

Grid stability under fault conditions is the principal constraint on the level of wind generation a network can accommodate. This work couples a Takagi–Sugeno–Kang fuzzy conditional inference engine to reinforcement learning, so that the controller adapts its own rule consequents rather than relying on a fixed rule base. It sustains a doubly-fed induction generator through voltage sags severe enough to defeat conventional ride-through schemes.

Adaptive fuzzy attention control of a microgrid under grid-bus fault

Double Q-learning · prioritized rewards · attention-weighted inference|Published · IEEE Trans. Fuzzy Systems 2025

A double Q-learning scheme with prioritized reward drives an attention-weighted fuzzy inference controller, keeping a microgrid stable through an extreme fault on the grid bus. This is the regime in which fixed-gain controllers fail, because the operating point departs further from nominal than their tuning assumes.

Photovoltaic modeling: from cell physics to systems

Conducted with Dr. A. H. M. A. Rahim and Prof. Md. Mosaddequr Rahman between 2023 and 2026. A custom-built bifacial module was characterized experimentally and then modeled from the one-diode equations upward.

Weather-responsive efficiency model for a custom-built bifacial panel

One-diode model · air-pressure and humidity terms · multiple cell technologies|Published · IEEE J. Photovoltaics 2024

Standard efficiency-rating models account only for irradiance and ambient temperature. Starting from the derivation of the one-diode model toward a photovoltaic efficiency rating, this work introduces air pressure and humidity as additional, carefully constructed dimensions, and validates the resulting model against modules based on several different cell technologies, including a bifacial panel constructed and characterized in-house.

Mono- and bifacial photovoltaic technologies compared

TOPCon · silicon heterojunction · next-generation bifacial cells|Published · IEEE J. Photovoltaics 2024

Next-generation bifacial cells are central to the development of high-efficiency modules. This review compares the emerging technologies, including TOPCon and silicon heterojunction, with their monofacial equivalents on a common basis.

Agrivoltaics: challenges and prospects

Dual land use · standards · community acceptance · policy|Published · Adv. Energy Sustain. Res. 2026

Agri-photovoltaics enables the dual use of land for agriculture and electricity generation. This review surveys recent agri-PV prospects across continents, together with the standards, community-acceptance and policy questions that determine whether the economic benefit reaches the farmers working beneath the arrays.