Spencer H. Bryngelson
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Spencer H. Bryngelson has a total of 113 co-authors. Here are the top co-authors by number of shared publications:
Tim Colonius
11 shared publications- Benjamin Wilfong
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA
7 shared publications - Anand Radhakrishnan
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA
6 shared publications Anand Radhakrishnan
5 shared publications- Florian Schaefer
Materials Science and Methods, Saarland University 1 , Saarbrücken
5 shared publications - Henry Le Berre
School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, GA, USA
5 shared publications
Latest Publications
- 31 October 2026
- Published by Elsevier in International Journal of Multiphase Flow
- Vol. 205
A sub-grid-scale model for polydisperse bubbly flows with heat and mass transfer
- 31 October 2026
- Published by Elsevier in International Journal of Multiphase Flow
- Vol. 205
A sub-grid-scale model for polydisperse bubbly flows with heat and mass transfer
- Open Access
- 29 October 2026
- Published by Wiley in Software: Practice and Experience
ANEForge: Python for Direct Computation on the Apple Neural Engine
Background: The Apple Neural Engine (ANE) is the fixed‐function neural accelerator on every recent Apple device, and the only sanctioned route to it is CoreML. CoreML treats the engine as a scheduling option: no configuration requires it, the CPU remains a permitted fallback, and no runtime interface reports which unit ran a call. Objective: We set out to program the ANE directly from Python, with the compute unit fixed so that each measurement belongs to the engine, and to report what the engine computes and at what cost. Methods: ANEForge compiles a lazy tensor graph into one ANE program and dispatches it through Espresso, the private runtime that CoreML is built on. It needs no entitlement and leaves system‐integrity protection on. The package reaches the engine's native fused‐attention layer, streams int8, int4, and sparse weights, keeps decoder and optimizer state on the engine across calls, runs the forward pass, backward pass, and optimizer update of training on the engine, and compiles fixed‐iteration Krylov solvers and a factored Fourier transform as static‐dataflow programs. We measured latency, energy per inference, and agreement with framework references on an M5 Pro, against the GPU and CPU at matched precision. Results: A pretrained ResNet‐18 forward runs in 0.33 ms and uses 2.3 mJ per inference, about a factor of 9 less than the GPU at matched fp16 precision (paired runs give 5.5 to 12.7). A call costs about 74 μs whether the program fuses one operator or 128. ResNet‐18, a Vision Transformer, and a sentence encoder each compile to one program and match their framework references at cosine similarity 1.0000. One forward pass of the Stable Diffusion U‐Net and VAE decoder agrees with the reference to a relative error of 0.041 and 0.021. Conclusions: The engine's advantage comes from fusion. A single memory‐bound operator finishes sooner on the CPU or the GPU, while a fused program pays the per‐call floor once. Fixed‐iteration methods whose program size does not depend on the problem dimension compile to the engine, and half precision limits how ill‐conditioned a problem they can solve.
- 29 October 2026
- Published by Wiley in Software: Practice and Experience
Open AccessANEForge: Python for Direct Computation on the Apple Neural Engine
Background: The Apple Neural Engine (ANE) is the fixed‐function neural accelerator on every recent Apple device, and the only sanctioned route to it is CoreML. CoreML treats the engine as a scheduling option: no configuration requires it, the CPU remains a permitted fallback, and no runtime interface reports which unit ran a call. Objective: We set out to program the ANE directly from Python, with the compute unit fixed so that each measurement belongs to the engine, and to report what the engine computes and at what cost. Methods: ANEForge compiles a lazy tensor graph into one ANE program and dispatches it through Espresso, the private runtime that CoreML is built on. It needs no entitlement and leaves system‐integrity protection on. The package reaches the engine's native fused‐attention layer, streams int8, int4, and sparse weights, keeps decoder and optimizer state on the engine across calls, runs the forward pass, backward pass, and optimizer update of training on the engine, and compiles fixed‐iteration Krylov solvers and a factored Fourier transform as static‐dataflow programs. We measured latency, energy per inference, and agreement with framework references on an M5 Pro, against the GPU and CPU at matched precision. Results: A pretrained ResNet‐18 forward runs in 0.33 ms and uses 2.3 mJ per inference, about a factor of 9 less than the GPU at matched fp16 precision (paired runs give 5.5 to 12.7). A call costs about 74 μs whether the program fuses one operator or 128. ResNet‐18, a Vision Transformer, and a sentence encoder each compile to one program and match their framework references at cosine similarity 1.0000. One forward pass of the Stable Diffusion U‐Net and VAE decoder agrees with the reference to a relative error of 0.041 and 0.021. Conclusions: The engine's advantage comes from fusion. A single memory‐bound operator finishes sooner on the CPU or the GPU, while a fused program pays the per‐call floor once. Fixed‐iteration methods whose program size does not depend on the problem dimension compile to the engine, and half precision limits how ill‐conditioned a problem they can solve.
- Open Access
- 10 August 2026
- Published by Cambridge University Press (CUP) in Journal of Fluid Mechanics
- Vol. 1040
Energy dissipation mechanisms in an acoustically driven slit
We quantify the conversion of incident acoustic energy into vortical motion and viscous dissipation for a plane wave passing through large-aspect-ratio slit geometries. We perform direct numerical simulations over a broad imposed parameter space in incident sound pressure level (ISPL), Strouhal number ( italic St $\textit{St}$ ) and Reynolds number ( italic Re $\textit{Re}$ ). Spectral proper orthogonal decomposition yields energy-ranked coherent structures at each frequency, from which we construct mode-by-mode fields for spectral kinetic energy (KE) and viscous loss (VL) to examine the acoustic absorption mechanisms. At italic ISPL equals 150 dB $\textit{ISPL} = {150}\,\textrm{dB}$ , the acoustic–hydrodynamic energy conversion is highest when italic St less than or equals 4 italic St Subscript 0 Baseline $\textit{St} \le 4\textit{St}_0$ , corresponding to an effective Keulegan–Carpenter number ( upper K Subscript c ${K_c}$ ) larger than 40. In this regime, three-dimensional simulations show that the dominant flow response is two-dimensional, with the oscillatory shear layer near the slit corners rolling up into shedding vortices. The VL accounts for 20 %–60 % of the KE contribution. For larger italic St $\textit{St}$ , the Stokes-layer confinement produces X-shaped near-slit modes, reducing the energy input by approximately 50 %. The influence of italic Re $\textit{Re}$ depends on amplitude. Larger italic Re $\textit{Re}$ corresponds to suppressed broadband fluctuations and sharpened harmonic peaks at 150 dB ${150}\,\textrm{dB}$ . At italic ISPL equals 120 dB $\textit{ISPL} = {120}\,\textrm{dB}$ , the boundary layers remain attached, vortex shedding is weak, absorption monotonically scales with viscosity and the italic Re $\textit{Re}$ - and italic St $\textit{St}$ -dependencies become comparable. Across all conditions, more than 99 % of the VL is confined to a compact region surrounding the slit mouth. The KE–VL spectra identify regimes that enhance or suppress acoustic damping in slit geometries, providing a physically interpretable basis for acoustic-based design.- 10 August 2026
- Published by Cambridge University Press (CUP) in Journal of Fluid Mechanics
- Vol. 1040
Open AccessEnergy dissipation mechanisms in an acoustically driven slit
We quantify the conversion of incident acoustic energy into vortical motion and viscous dissipation for a plane wave passing through large-aspect-ratio slit geometries. We perform direct numerical simulations over a broad imposed parameter space in incident sound pressure level (ISPL), Strouhal number ( italic St $\textit{St}$ ) and Reynolds number ( italic Re $\textit{Re}$ ). Spectral proper orthogonal decomposition yields energy-ranked coherent structures at each frequency, from which we construct mode-by-mode fields for spectral kinetic energy (KE) and viscous loss (VL) to examine the acoustic absorption mechanisms. At italic ISPL equals 150 dB $\textit{ISPL} = {150}\,\textrm{dB}$ , the acoustic–hydrodynamic energy conversion is highest when italic St less than or equals 4 italic St Subscript 0 Baseline $\textit{St} \le 4\textit{St}_0$ , corresponding to an effective Keulegan–Carpenter number ( upper K Subscript c ${K_c}$ ) larger than 40. In this regime, three-dimensional simulations show that the dominant flow response is two-dimensional, with the oscillatory shear layer near the slit corners rolling up into shedding vortices. The VL accounts for 20 %–60 % of the KE contribution. For larger italic St $\textit{St}$ , the Stokes-layer confinement produces X-shaped near-slit modes, reducing the energy input by approximately 50 %. The influence of italic Re $\textit{Re}$ depends on amplitude. Larger italic Re $\textit{Re}$ corresponds to suppressed broadband fluctuations and sharpened harmonic peaks at 150 dB ${150}\,\textrm{dB}$ . At italic ISPL equals 120 dB $\textit{ISPL} = {120}\,\textrm{dB}$ , the boundary layers remain attached, vortex shedding is weak, absorption monotonically scales with viscosity and the italic Re $\textit{Re}$ - and italic St $\textit{St}$ -dependencies become comparable. Across all conditions, more than 99 % of the VL is confined to a compact region surrounding the slit mouth. The KE–VL spectra identify regimes that enhance or suppress acoustic damping in slit geometries, providing a physically interpretable basis for acoustic-based design. - Open Access
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
A sub-grid-scale model for polydisperse bubbly flows with heat and mass transfer
Ensemble-averaged models of polydisperse bubbly flows require statistics of the evolving bubble population. Prior quadrature-based moment formulations close bubble pressure with a polytropic relation that omits heat and mass transfer at the bubble wall. We formulate constant-transfer equations for bubble pressure and vapor mass within a conditional hyperbolic quadrature method. Second-order conditional inversion produces four joint radius–radial-velocity nodes per equilibrium-radius bin. Bubble pressure and vapor mass are advanced at each node. The node values close the ensemble-averaged flow equations without adding mixed pressure or vapor-mass moments to the transported moment set. The model is implemented in MFC. Monte Carlo calculations verify the evolution of quadrature nodes and the mean bubble variables for a harmonically forced population. Bubble-screen calculations quantify closure error as the equilibrium radius is discretized and the initial distributions vary. The constant-transfer calculation does not exhibit the high-frequency pressure oscillations observed with the polytropic closure under the conditions considered. 3D bubble-screen calculations give a 1.5% relative root-mean-square error between the Euler–Euler center pressure and the mean of 40 volume-averaged Euler–Lagrange realizations.
- 31 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA sub-grid-scale model for polydisperse bubbly flows with heat and mass transfer
Ensemble-averaged models of polydisperse bubbly flows require statistics of the evolving bubble population. Prior quadrature-based moment formulations close bubble pressure with a polytropic relation that omits heat and mass transfer at the bubble wall. We formulate constant-transfer equations for bubble pressure and vapor mass within a conditional hyperbolic quadrature method. Second-order conditional inversion produces four joint radius–radial-velocity nodes per equilibrium-radius bin. Bubble pressure and vapor mass are advanced at each node. The node values close the ensemble-averaged flow equations without adding mixed pressure or vapor-mass moments to the transported moment set. The model is implemented in MFC. Monte Carlo calculations verify the evolution of quadrature nodes and the mean bubble variables for a harmonically forced population. Bubble-screen calculations quantify closure error as the equilibrium radius is discretized and the initial distributions vary. The constant-transfer calculation does not exhibit the high-frequency pressure oscillations observed with the polytropic closure under the conditions considered. 3D bubble-screen calculations give a 1.5% relative root-mean-square error between the Euler–Euler center pressure and the mean of 40 volume-averaged Euler–Lagrange realizations.
- Open Access
- 1 July 2026
- Published by Springer Nature in Communications Physics
- Vol. 9 (1)
Hadamard random forest for reconstructing real valued pure quantum states
- 1 July 2026
- Published by Springer Nature in Communications Physics
- Vol. 9 (1)
Open AccessHadamard random forest for reconstructing real valued pure quantum states
- Open Access
- 30 June 2026
- Published by IOP Publishing in Quantum Science and Technology
- Vol. 11 (3) , 035032
Quantum lattice Boltzmann algorithm for heat transfer with phase change
Heat transfer involving phase change is computationally intensive due to moving phase boundaries, nonlinear computations, and time step restrictions. This paper presents a quantum lattice Boltzmann method (QLBM) for simulating heat transfer with phase change. The approach leverages the statistical nature of the lattice Boltzmann method (LBM) while addressing the challenges of discontinuous phase transitions in quantum computing. The method implements an interface-tracking strategy that partitions the problem into separate solid and liquid domains, enabling the algorithm to handle the discontinuity in the enthalpy–temperature relationship. We store phase change information in the quantum circuit to reduce information exchange between classical and quantum hardware, a bottleneck in many quantum applications. Results from the implementation agree with both classical LBM and analytical solutions, demonstrating QLBM as an effective approach for analyzing thermal systems with phase transitions. Simulations using 17 lattice nodes with 53 qubits demonstrate temperature root-mean-square errors of order 0.01 when compared against classical solutions. The method accurately tracks interface movement during phase transition.
- 30 June 2026
- Published by IOP Publishing in Quantum Science and Technology
- Vol. 11 (3) , 035032
Open AccessQuantum lattice Boltzmann algorithm for heat transfer with phase change
Heat transfer involving phase change is computationally intensive due to moving phase boundaries, nonlinear computations, and time step restrictions. This paper presents a quantum lattice Boltzmann method (QLBM) for simulating heat transfer with phase change. The approach leverages the statistical nature of the lattice Boltzmann method (LBM) while addressing the challenges of discontinuous phase transitions in quantum computing. The method implements an interface-tracking strategy that partitions the problem into separate solid and liquid domains, enabling the algorithm to handle the discontinuity in the enthalpy–temperature relationship. We store phase change information in the quantum circuit to reduce information exchange between classical and quantum hardware, a bottleneck in many quantum applications. Results from the implementation agree with both classical LBM and analytical solutions, demonstrating QLBM as an effective approach for analyzing thermal systems with phase transitions. Simulations using 17 lattice nodes with 53 qubits demonstrate temperature root-mean-square errors of order 0.01 when compared against classical solutions. The method accurately tracks interface movement during phase transition.
- Open Access
- 23 April 2026
- Published by Elsevier in SSRN Electronic Journal
Shocks without shock capturing: Information geometric regularization of finite volume methods for Navier–Stokes-like problems
Shock waves in high-speed fluid dynamics produce near-discontinuities in the fluid momentum, density, and energy. Most contemporary works use artificial viscosity or limiters as numerical mitigation of the Gibbs–Runge oscillations that result from traditional numerics. These approaches face a delicate balance in achieving sufficiently regular solutions without dissipating fine-scale features, such as turbulence or acoustics. Recent work by Cao and Schäfer introduces information geometric regularization (IGR), the first inviscid regularization method for fluid dynamics. IGR replaces shock singularities with smooth profiles of adjustable width, without dissipating fine-scale features. This work provides a strategy for the practical use of IGR in finite-volume-based numerical methods. We illustrate its performance on canonical test problems and compare it against established approaches based on limiters and Riemann solvers. Results show that the finite volume IGR approach recovers the expected solutions in all cases. Across canonical benchmarks, IGR achieves accuracy competitive with WENO and LAD shock-capturing schemes in both smooth and discontinuous flow regimes. The IGR approach is computationally light, with meaningfully fewer memory accesses and arithmetic operations per time step.
- 23 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessShocks without shock capturing: Information geometric regularization of finite volume methods for Navier–Stokes-like problems
Shock waves in high-speed fluid dynamics produce near-discontinuities in the fluid momentum, density, and energy. Most contemporary works use artificial viscosity or limiters as numerical mitigation of the Gibbs–Runge oscillations that result from traditional numerics. These approaches face a delicate balance in achieving sufficiently regular solutions without dissipating fine-scale features, such as turbulence or acoustics. Recent work by Cao and Schäfer introduces information geometric regularization (IGR), the first inviscid regularization method for fluid dynamics. IGR replaces shock singularities with smooth profiles of adjustable width, without dissipating fine-scale features. This work provides a strategy for the practical use of IGR in finite-volume-based numerical methods. We illustrate its performance on canonical test problems and compare it against established approaches based on limiters and Riemann solvers. Results show that the finite volume IGR approach recovers the expected solutions in all cases. Across canonical benchmarks, IGR achieves accuracy competitive with WENO and LAD shock-capturing schemes in both smooth and discontinuous flow regimes. The IGR approach is computationally light, with meaningfully fewer memory accesses and arithmetic operations per time step.
- Open Access
- 21 March 2026
- Published by Elsevier in International Journal of Multiphase Flow
- Vol. 199
Hardware-accelerated phase-averaging for cavitating bubbly flows
- 21 March 2026
- Published by Elsevier in International Journal of Multiphase Flow
- Vol. 199
Open AccessHardware-accelerated phase-averaging for cavitating bubbly flows
- Open Access
- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
Fourth-Order HyQMOM Closures for Multidimensional Kinetic Equations
The hyperbolic quadrature method of moments (HyQMOM) for closing the moment system found from a one-dimensional (1D) kinetic equation was introduced in prior work (Fox \& Laurent, SIAM J.\ Appl.\ Math., 82 (2022)). Here, HyQMOM is extended to two-dimensional (2D) kinetic equations using up to 15 fourth-order velocity moments and three-dimensional (3D) kinetic equations with 35 moments. Using the 15-moment system, example calculations are carried out in a 2D spatial domain for Riemann problems involving impinging jets for various values of the Knudsen ($\Kn$) and jet Mach ($\Ma$) numbers. Due to the structure of the 2D HyQMOM closure for the fifth-order moments and the correction algorithm for moment realizability and hyperbolicity, a robust numerical algorithm based on the HLL solver can be applied to simulate cases with arbitrary values of $\Kn$ and $\Ma$. The extension to 3D kinetic equations requires twenty additional moments, which can be derived and implemented numerically relatively straightforwardly. The principal implementation difficulty in 3D is the formulation of the moment-correction algorithm. Nonetheless, the 35-moment system is found to be hyperbolic under the same conditions as the 15-moment system. Simulations of the 3D kinetic equation in 2D and 3D spatial domains demonstrate the robustness of the closure for arbitrary $\Kn$ and $\Ma$.- 9 February 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFourth-Order HyQMOM Closures for Multidimensional Kinetic Equations
The hyperbolic quadrature method of moments (HyQMOM) for closing the moment system found from a one-dimensional (1D) kinetic equation was introduced in prior work (Fox \& Laurent, SIAM J.\ Appl.\ Math., 82 (2022)). Here, HyQMOM is extended to two-dimensional (2D) kinetic equations using up to 15 fourth-order velocity moments and three-dimensional (3D) kinetic equations with 35 moments. Using the 15-moment system, example calculations are carried out in a 2D spatial domain for Riemann problems involving impinging jets for various values of the Knudsen ($\Kn$) and jet Mach ($\Ma$) numbers. Due to the structure of the 2D HyQMOM closure for the fifth-order moments and the correction algorithm for moment realizability and hyperbolicity, a robust numerical algorithm based on the HLL solver can be applied to simulate cases with arbitrary values of $\Kn$ and $\Ma$. The extension to 3D kinetic equations requires twenty additional moments, which can be derived and implemented numerically relatively straightforwardly. The principal implementation difficulty in 3D is the formulation of the moment-correction algorithm. Nonetheless, the 35-moment system is found to be hyperbolic under the same conditions as the 15-moment system. Simulations of the 3D kinetic equation in 2D and 3D spatial domains demonstrate the robustness of the closure for arbitrary $\Kn$ and $\Ma$.