Research
Device physics carried through the instrument, so that the model ends at the number a measurement returns.
Each project below is built as a single description that runs from a material’s electronic structure to the quantity an instrument reports: nothing is fitted where it can be computed. And because the model’s sensitivities can be traced end to end, the chain also runs in reverse: to recover a hidden quantity from data, or to turn a target specification into a geometry.
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.
Squeezed-light microcombs, diamond quantum sensing and photonic machine learning
Conducted with Dr. A. S. M. Mohsin, Department of EEE, BRAC University, since January 2025. This work develops models of quantum photonic hardware, together with learned controllers that close the loop around it: squeezed-light sources, nitrogen-vacancy (NV) diamond magnetometers and single-photon instruments.
Overcoming the 3 dB squeezing extraction limit in silicon-carbide microcombs
Squeezed light is light whose noise, in the right measurement, falls below the usual quantum limit, a resource for precision measurement and quantum computing. This work builds an open, reproducible pipeline for producing strongly squeezed light in microcombs (micro-rings that turn one pump laser into many evenly spaced colors) on 4H-silicon-carbide-on-insulator. Modeling of the waveguide, simulation of the comb dynamics (Lugiato–Lefever) and a linearized quantum-noise analysis are chained into one workflow, running from material parameters to the noise a detector would record. A single ring faces a built-in trade-off: the coupling that builds up light inside the ring and the coupling that lets the squeezing out oppose each other, so detectable squeezing stalls at 3 dB no matter how hard the ring is pumped. The fix is a second, auxiliary ring, with resonances spaced twice as far apart, which pulls the squeezed modes out while leaving the classical comb untouched. At a fixed 8.3 mW pump, the best extraction rate lies near ten cavity linewidths, and the full two-ring model predicts 7.9 dB of detectable squeezing over 1.81 GHz, rising to 8.5 dB near the boundary of comb stability, and concentrated in two dominant squeezed supermodes with an entangled odd-mode lattice. One more result: third-order dispersion shifts the comb's repetition rate measurably, yet leaves the squeezing unchanged right up to the point where it destroys the soliton crystal. Together with the explicit two-ring design, that answers the two open problems the original photonic-molecule proposal left standing: how much higher-order dispersion the squeezing tolerates, and the design of the molecule itself.
Fast single-shot readout of tin-vacancy spins on thin-film lithium niobate
A tin-vacancy center is a defect in diamond that stores a quantum bit in its spin and talks to light. That makes it a candidate node for quantum networks. But it has a built-in conflict. Strain in the crystal makes microwave control of the spin fast. The same strain makes the spin flip more often while it is being read. Single-shot readout needs many detected photons before such a flip, and a microscope detects only 0.1 to 0.4% of them. So readout has limited both the strain and the direction of the magnetic field.
This work shows by simulation that collecting the light efficiently removes the conflict. A diamond nanobeam cavity with an intrinsic quality factor of 1.4 × 106 is deliberately spoiled. It is opened on one side to a loaded quality factor of 870 (Purcell factor 28), and 92% of the light it releases enters the output waveguide. A short taper then passes 99.7% of that light to a lithium niobate waveguide on the same chip. With measured fiber-coupling and detector efficiencies, 32% of the emitted photons are detected, about a hundred times more than with a microscope. A four-level model, built from the strained spin's own energy levels, predicts 98.8% single-shot fidelity in 0.1 µs, against a measured 87.4% in 50 µs with a microscope. The fidelity stays above 98.1% for every strain studied, so strain can be chosen for fast control. It stays above 90% for magnetic fields up to 40° away from the spin axis.
Why spoil a good cavity? Once its Purcell factor passes about ten, a stronger cavity collects only a little more light. But it also broadens the optical line, and the broader line starts to excite the dark spin through its own transition, only 0.73 GHz away. Those stray photons blur the difference between the two spin states. So the fidelity is best for a moderate, strongly overcoupled cavity. The design also gives a practical rule for the taper. A 600 nm lithium niobate waveguide merges the diamond and lithium niobate modes smoothly, while a 400 nm one passes almost no light. The spin model reproduces published measurements of the same emitter with no adjusted parameters. The taper, the cavity loading and the readout are checked against errors in placement, etching, linewidth and background light. All results are simulated. The largest risk not modeled is that the center's charge state becomes unstable under the readout laser. The general-purpose tool vacspin, on the Software page, grew out of this study.
Listening to vortex noise: which trapped vortices cause loss in tantalum qubits
Vortices are tiny whirlpools of magnetic flux. They get trapped when a superconducting film is cooled in a stray magnetic field. When a microwave current pushes a vortex, its core moves and wastes energy. This loss has been seen in tantalum, the film used in record-coherence qubits, and patterned holes reduce it. But a resonator measurement gives only the total loss. It cannot say which trapped flux is lossy. A static magnetic image cannot say either: flux held in a hole and a vortex between holes look alike from above.
This work shows that the noise of a vortex answers the question, and also measures its loss. A basic law of physics, the fluctuation–dissipation theorem, says that anything that absorbs energy when driven must also jitter on its own. So a lossy vortex is noisy, and flux held in a hole is silent. A diamond spin sensor (a nitrogen-vacancy, or NV, centre) can hear this noise. The calculation uses measured tantalum values, plus an assumed film thickness and penetration depth, which are scanned. It finds that one moving vortex makes an NV spin 50 nm away relax at 1.4 × 105 per second at 2 K, about 6,000 times the rate caused by the superconducting film's own noise. Simulations of cooled films confirm the picture. Flux held in holes is as bright as a vortex in a static image, but adds no noise. Under a 5 GHz drive, each vortex between holes wastes energy within 6% of the textbook value for a freely moving vortex, while flux held in a hole wastes at most 9% as much.
For vortices held only weakly in place, the noise therefore works as a calibrated loss meter, with no free parameter. In a test on noisy simulated data, it recovered the loss of each vortex within 1.5% whenever the vortex's pinning frequency was below about 2 GHz. It also gives a design rule. Stronger pinning lowers the loss only if it pushes the pinning frequency above the qubit frequency. For a 5 GHz qubit on tantalum, that needs a pinning stiffness above 2.3 × 10−4 N/m. Holes work differently, because they hold flux without a core. In simulated 2 × 2 µm films at one cooling field, holes cut the number of lossy vortices from 19 to 2 or 3, although the count also depends on how the film is cooled. The proposed experiment needs only parts that already exist. It would count the lossy vortices at 2 K and predict their loss before the qubit is cooled to millikelvin temperatures, assuming the vortex drag does not change with temperature.
SPARQ: autonomous triage of solid-state single-photon emitters
A single-photon emitter gives out light one photon at a time, and many quantum technologies need one. A sample holds many candidate spots, and only a minority are single emitters. The standard test, the Hanbury Brown–Twiss measurement, splits the light onto two detectors and records the time gaps between photons arriving at each. It is slow, because the useful photon pairs build up at a rate set by the square of the brightness. This work asks how much of that time can be saved.
Three parts work together. A spiking neural network reads the photon data in short slices and updates its verdict after each one, so it can decide early. A reinforcement-learning controller chooses, spot by spot, whether to measure longer, certify, or reject. A small graph of each emitter type's energy levels lets one estimator handle several types. All parts are trained on simulated data from a standard emitter model, which is first checked against the exact solution.
In simulation, the spiking network reaches the 90% accuracy target with 6 times less measuring time than a standard least-squares fit. It can also decide early: typically 312 ms into a 1 s exposure, at 86% accuracy. In the closed-loop tests, a standard (non-spiking) network supplied the estimates. Adapting the time spent on each spot then screens simulated 48-spot fields 1.7 times faster than a fixed scan, at the same quality. On four emitter types it never saw in training, the energy-level graphs close most of the accuracy gap to a model trained on those types, also in simulation. On real published data from a quantum dot, a network trained only on simulated data estimated the purity value g(2)(0) from 30 s records with half the error of the usual peak-area analysis.
The paper also reports what did not help. A simple rule that stops once the estimate is confident was faster still: 3.8 times faster than the fixed scan, against 1.7 times for the learned controller. The spiking network showed no energy advantage. The closed-loop results are in simulation only.
PILOT-Q: photon-efficient neural inference at the standard quantum limit
In delocalized photonic inference, a central server sends a neural network's weights as light to small edge devices, which do the multiplications optically. These links work at about one photon per multiplication. There, shot noise, the unavoidable randomness of counting photons, controls the result. Yet such links run models trained with clean digital arithmetic. They also give every input the same photon budget, sized for the hardest one.
PILOT-Q builds a noise model of the receiver that a network can be trained through. It covers shot noise at the quantum limit, dark counts, weight errors and digitizing. It matches the textbook shot-noise law within 9% over more than two orders of magnitude in photon budget. In simulation, training through it cuts the photons needed for the same accuracy by 1.4 to 3.1 times on three open benchmarks. With all impairments on, at one photon per multiplication, accuracy on spoken digits rises from 51.6% to 78.6%. Clean accuracy drops by at most 1.7 points.
Two results are negative, and useful. Adding the other hardware faults during training made results worse: shot noise alone gives the robustness. And a controller that re-measures only uncertain inputs adds little. On handwritten digits (MNIST), even a perfect controller that re-measures once, and knows which inputs will be wrong, could save only about 2 times. At the quantum limit, training matters more than run-time tricks. The photon savings lower the server's laser power and the measuring time; the edge device's own energy is set by its electronics.
FabGAN-ID: learned fabrication twins for yield-aware photonic design
A multilayer optical filter is a stack of thin layers, and no deposition machine builds exactly the stack it is asked for. Designs that must survive this are usually optimized against an assumed error model, most often independent Gaussian (bell-shaped) errors. Real deposition errors are different. They are systematic, correlated between layers, dependent on the design, skewed and heavy-tailed, and the manufacturing yield is decided in that tail.
FabGAN-ID replaces the assumed error model with a learned one: a conditional, moment-matched Wasserstein GAN trained on only 400 process traces of a variable-index SiNx multilayer platform. On the statistics that decide yield (the 5th percentile and the CVaR, the average over the worst 5%), its error is 34 to 38% lower than that of the best of eight standard density models fitted to the same data. Because the optical solver and the GAN are both differentiable, the design is optimized through both at once. On a held-out true process, this raises the lower-tail merit of a 532 nm fluorescence-rejection notch filter by 7.2% over the nominal design. That matches an independent oracle that knows the true process, and beats every robust design built on a Gaussian error model. The simulated pass/fail yield rises from 0.16% to 63.7%.
The study is fully simulation-based: the traces come from a virtual process, and the next step is to learn the twin from real wafer data. In an exploratory comparison, the same gradients also train correction policies, where model-free reinforcement learning did not converge. Neural surrogates of the optics failed on held-out designs and ran 18 times slower than the exact solver. The benchmark (300 spectra, 48,300 wavelength and transmittance samples, and 400 traces) is released openly.
GaN and 2D semiconductor devices and physics-guided design
Conducted with Prof. Md. Mosaddequr Rahman, Department of EEE, BRAC University, from July 2024 to July 2026. This work follows charge carriers and lattice vibrations (phonons) through nitride and single-layer semiconductor channels, from the underlying quantum states up to the masses, lifetimes and circuit-level numbers experiments report, and applies physics-in-the-loop inverse design to micromachined transducers. A collaboration with Dr. Nadim Chowdhury at BUET carries the same physics-guided approach into circuit design, described at the end of this section.
Origin of the conflicting hole masses in the GaN/AlN two-dimensional hole gas
A two-dimensional hole gas is a thin sheet of positive charge carriers (holes) trapped at an interface. The one at a GaN/AlN interface could give GaN the p-type transistors that GaN logic circuits need. Three experiments have measured it: quantum oscillations in magnetic fields up to 72 T, terahertz cyclotron resonance up to 31 T, and a two-carrier analysis of the Hall effect. Their results disagree. The heavy-hole mass is 1.92 m0 in one and 2.6 m0 in another (m0 is the mass of a free electron). The light-hole mass, 0.53 m0, is about twice what theory gives. And the scattering behind the measured mobilities has not been pinned down.
This work solves the standard six-band model of the valence band together with the electrostatics, at the measured hole density, using published parameters. The heavy-hole mass comes out at 1.92 to 1.99 m0, matching the measured 1.92 ± 0.16 m0 with nothing adjusted. The cyclotron value differs because that resonance is overdamped: the heavy holes scatter before they finish enough of an orbit (ωcτ = 0.82 at 31 T, below the value of one a clean resonance needs).
The light holes carry one real discrepancy. The measured densities and both measured light-hole masses need a light-hole band about one and a half times heavier than calculated. Of all the band parameters, only one, called A6, can do this without disturbing the heavy holes. Rescaled to fit, it also matches the light-hole mass measured on a second sample, and it gives a mass that rises with magnetic field, by up to 58% of the reported rise. The value it needs is at or slightly beyond the limit the model allows, so it is treated as an effective parameter. An independent measurement of A6 would test it.
For scattering, no single mechanism explains both subbands at once. The heavy-hole quantum mobility had been estimated from the field where its oscillations first appear. That overestimates it, because the heavy holes carry most of the current. A fit of the published oscillations gives 95 cm2 V−1 s−1 instead of 167 to 200. With this value, the four measured mobilities fix the kind of disorder. On top of interface roughness, there is smooth, long-range disorder that needs charged line defects threading the hole gas, together with a flatter component. The two-carrier Hall analysis is not the cause of the disagreement.
Two Raman peaks and transistor thresholds measure the fixed charge at etched MoS2 edges
Cutting a single-layer semiconductor into a narrow ribbon can leave electric charge or mechanical strain at its edges. Either one can limit how narrow a working transistor can be. Both shift Raman peaks, so one peak cannot tell them apart. To our knowledge, no edge charge density had been reported for such a ribbon.
Two peaks of the same spectrum can tell them apart. First-principles calculations on MoS2, WS2, MoSe2 and WSe2 show that the 2LA(M) band is several times more sensitive to strain than the A′₁ peak (4.1 times in MoS2). A direct measurement on WSe2 suggests the calculation overstates this, but the edge result does not depend on it. In MoS2, the A′₁ peak also responds strongly to added electrons. Applied to published tip-enhanced Raman maps of an etched MoS2 ribbon, the pair shows that its edges hold extra electrons, 2.3 × 1012 ± 7.6 × 1011 per cm2, with strain below 0.03%. A spot inside the ribbon is strained (0.134% tension) but not charged.
Published threshold voltages of 37 MoS2 transistors, made by the same etch, give a second and independent measure. A model that solves the electric field and the electrons together shows that the fixed edge charge is negative, reaches about 12 nm into the ribbon, and amounts to 4.3 × 107 per cm of edge. Ribbons cut by another etch on a thick oxide give the same order. This is 19 times the Raman value. The paper reads the Raman value as the part of the charge seen as mobile electrons in a sample measured on gold.
The rest is model prediction. On a thin high-κ gate, with an assumed 5 nm damaged strip, the critical width is 17 nm with the Raman charge and 96 nm with the transistor-derived charge. Most of that rise is a threshold shift, which gate design can offset: compared at the same drive above each ribbon's own threshold, the critical width is 26 nm. With nothing fitted, the model's currents for two published n-type ribbon transistors lie 3% and 19% above the measured ones. A p-type one needs a contact barrier of 0.26 eV, within the range computed for such contacts. Film quality changes how much current a ribbon carries, but barely moves its critical width.
What sets the memory window of a 2D ferroelectric transistor
A ferroelectric transistor stores a bit in the direction of a ferroelectric layer's polarization. Its memory window is the gap between its two threshold voltages, one for each stored state. CIPS is a layered ferroelectric that has been stacked on several two-dimensional channels. A natural question is which properties of the channel can change the window.
This work derives a simple formula for the window, read at a fixed current. It holds when the internal floating gate is an ideal conductor, the ferroelectric carries no extra charge, and the current depends on the channel in the same way on both sweeps. Then the channel's own properties cancel out. The window depends only on the ferroelectric thickness, its switched polarization, and the charge trapped at the channel surface. In simulation the formula agrees with the full model to 0.02%, even while strain changes the hole mobility 26-fold. This checks the maths of the model, not the physics.
The channel still matters, but only in limited ways. It acts through the polarization left behind after programming. The window stays independent of the channel only when two things hold. The channel must take on charge about ten times more easily than the h-BN layer at full programming. And the read current must stay below the point where switching starts. On the channel side, trapped interface charge is the main limit: the window loses less than 6% up to about 1012 traps per cm2 per eV, a level that h-BN interfaces meet but CVD-grown channels do not.
In the model, strain therefore becomes a pure transport knob. At 1% compression, the hole mobility rises 2.38 times and the stored read current 1.90 times, while the window barely moves: 1.4342 V without strain and 1.4349 V at 1% compression. Projected circuits, with an assumed n-type partner transistor, keep their state through power loss, with a static noise margin of 642 mV. Their worst-case static power is 5.0 fW, against 1.8 µW if the same stage used the 5 MΩ load resistor of an earlier CIPS inverter. The work is theoretical: no device was made. It proposes a single fabrication run that could prove these predictions wrong.
PARL-ID: fabrication-aware inverse design across MEMS and photonics
PARL-ID unifies physics-informed neural networks, adjoint (gradient-based) optimization and reinforcement learning into a single framework for designing sensors that stay within specification when manufacturing varies. Validated on micromachined ultrasound transducers (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: that work identified the best network architecture for a nominal design; this one optimizes for designs that survive process variation.
Hierarchical inverse design of unit-cell CMUTs
Designing a micromachined ultrasound transducer normally means a manual sweep over device profiles. In this published work, a probabilistic search first identifies the machine-learning architecture best suited to the task (attentive gated recurrent layers feeding fully connected dense layers), and that network then maps a target acoustic response back to unit-cell CMUT geometry, maximizing membrane displacement without the exhaustive finite-element sweeps that unit-cell design normally demands.
Physics-guided circuit design. Conducted in collaboration with Dr. Nadim Chowdhury, Department of EEE, BUET, from May 2025 to June 2026. This work carries gradients through the compact models and the loop equations of a GaN circuit, so that the circuit can be designed by optimization, and develops reinforcement-learning methods that transfer across process nodes.
Device-physics design rules for wide-temperature GaN integrated circuits
GaN integrated-circuit processes offer only one kind of transistor, and its properties change with temperature. This study asks what that means for a complete circuit that must work from 218 to 423 K (−55 to 150 °C), the range power transistors are specified for. The test case is a monolithic GaN-on-SOI frequency synthesizer, a fractional-N phase-locked loop built only from these transistors. Learned models sized it: a physics-informed neural network model of the varactor (0.35 % error on held-out data) and a loop model with exact gradients, re-fitted to 520 transistor-level simulations. Checking the design with the foundry device model revealed three rules. First, bootstrapped logic loses its drive when hot: with the bootstrap charged from the supply, the ring oscillator stops at 398 K and above, and a precharge only 0.25 V above the supply restores it. Second, the varactor changes its capacitance only 3.3-fold, while the oscillator frequency changes four- to five-fold, so capacitance must be switched. Third, two kinds of on-chip resistor change differently with temperature, which gives a temperature sensor without p-type or bipolar devices; its weak point is the comparator, since a 0.2 V threshold shift moves its switching points by more than 30 K. The final design locks at 50 and 52 MHz at all ten temperatures in transistor-level simulation. Each rule is stated as a prediction that simple test structures can check.
Transferable analog transistor sizing with graph reinforcement learning
Developed with GlobalFoundries, Inc., this framework sizes analog transistors automatically. A graph network reads the circuit (devices as nodes, wires as edges), a reinforcement-learning agent proposes sizes, a fuzzy reward model carries the uncertainty in what "good" means, and physics-supplied gradients guide the search directly. Whereas conventional approaches rely on black-box circuit simulation and fixed weighted objectives, this combination improves both the quality of the result and the number of simulations needed to reach it. Validated across multiple technology nodes and amplifier benchmarks, it consistently outperforms existing Bayesian-optimization and reinforcement-learning methods while transferring across circuit topologies.
Learning-based grid control and bifacial photovoltaics
Conducted with Dr. A. H. M. A. Rahim (retired December 2024) from May 2023 to June 2024. This work develops controllers that adapt their own inference rules, applied to grid-connected machines under fault conditions, and a custom-built bifacial solar module.
Fuzzy inference with reinforcement learning for DFIG low-voltage ride-through
How much wind generation a grid can accept is limited by what happens during faults. This work couples a Takagi–Sugeno–Kang fuzzy inference engine (control rules expressed in graded, human-readable form) to reinforcement learning, so that the controller adapts its own rules rather than relying on a fixed rule base. It sustains a doubly-fed induction generator, the standard wind-turbine generator, through voltage sags severe enough to defeat conventional ride-through schemes.
Adaptive fuzzy attention control of a microgrid under grid-bus fault
A double Q-learning scheme with prioritized rewards 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.
Bifacial photovoltaics. Conducted with Dr. A. H. M. A. Rahim and Prof. Md. Mosaddequr Rahman. A custom-built bifacial module was characterized experimentally and then modeled from the one-diode equations upward. The agrivoltaics review that closes this section is a later publication, written with Prof. Md. Mosaddequr Rahman and Dr. A. S. Nazmul Huda.
Weather-responsive efficiency model for a custom-built bifacial panel
Standard efficiency-rating models account only for sunlight intensity 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
Next-generation bifacial cells (panels that collect light on both faces) 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
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.
Open-source software
Beyond the per-paper repositories above, twelve general-purpose tools distilled from this research (ramansep, kpenvelope, sqzcomb, absnoise, cavsqueeze, SPARQ, hamop, fabtwin, vacspin, labplan, fracpll and lockkernel) are maintained under the TaN-MM-Org organization, with tested cores, continuous integration and archived DOIs. They have a page of their own: Software.