Final Technical Report Collaborative Research: Enabling multi-scale studies of magnetic reconnection with interpretable data-driven models DOE award number: DE-SC0022012 Sponsoring program office: Office of Science Name of recipient: Massachusetts Institute of Technology Project title: Collaborative Research: Enabling multi-scale studies of magnetic reconnection with interpretable data-driven models Period of Performance: 09/01/2021 through 08/31/2025 Principal Investigator: Nuno Loureiro (Institutional PI at MIT) Collaborating Institutions: Stanford University: Frederico Fiuza (Lead PI) University of California, Los Angeles (UCLA) E. Paulo Alves (Institutional PI) University of Maryland William Dorland (Institutional PI) PI = Principal Investigator 1.0 Project goals The development of accurate reduced descriptions and improved closures for magnetic reconnection is an important and a long‐standing challenge in plasma physics. The four‐fluid approach, and associated closures, that were investigated have the potential to improve the accuracy of plasma fluid models, capturing physical effects which would otherwise require a kinetic description. If successful, this approach could have an important impact for the modeling of laboratory and space plasmas. The major goals of this project were to develop new machine learning (ML) tools based on sparse and symbolic regression techniques, and to extract interpretable and generalizable reduced models (e.g., in the form of partial differential equations - PDEs) from data generated by first principles plasma simulations. Preserving interpretability of such data‐driven models is key to addressing the long‐standing theoretical and numerical challenges. Prior proof‐of‐principle studies have demonstrated the enormous potential of this approach, by recovering the well‐ 1 established hierarchy of plasma equations (from Vlasov to MHD) from data produced by particle‐in‐cell (PIC) simulations. Our goal in this project was to extend and apply these new tools to construct better kinetic closures for magnetic reconnection; to derive better models of particle injection and acceleration by this fundamental plasma process; and to use this understanding to accelerate the development of multi‐scale plasma algorithms. This aspect of the project was carried out primarily by our collaborators at Stanford University (Frederico Fiuza) and at University of California, Los Angeles (E. Paulo Alves) [see Alves & Fiuza 2025]. While our immediate focus was on the problem of magnetic reconnection, the tools that were will developed are general and applicable to other areas of plasma physics, and more broadly to many‐body phenomena. We anticipate that the development of these multi‐scale models will have a significant impact across different areas of plasma science, from fusion to space and astrophysical plasmas. 2.0 Project Accomplishments 2.1 Reconnection closures The primary role of MIT in this project was to carry out first-principle simulation studies of magnetic reconnection processes with the goal of informing the machine-learning based development of reduced models and closure schemes. A careful analysis of some “reconnection closures” present in the literature was undertaken which has revealed their limitations. By comparing fully kinetic and two‐fluid simulations of magnetic reconnection in low‐β plasmas, we have shown counter‐streaming between electron populations can play an important role by modifying the electron pressure near the X‐point and separatrix regions. This effect is not captured by current two‐fluid simulations. We have shown that decomposing the electrons species into two separate fluids (i.e., an electron fluid that enters the reconnection region from above, and another one from below), despite its simplicity, can in fact capture the counter‐ streaming physics and reproduce the dynamical behavior seen in fully kinetic particle-in-cell simulations. In particular, the triggering of secondary instabilities in the current sheet and the rate of reconnected magnetic flux are well reproduced by the two‐electron‐fluid model. Interestingly, this separation of electron fluids enables the use of a simple adiabatic closure at the level of each individual fluid component, while capturing important modifications to the total electron pressures tensor and heat flux of the combined electron fluids. This approach thus not only provides a simple fluid closure that reproduces important dynamical behavior observed in fully kinetic simulations but also provides an entirely new perspective on the importance and role of counter‐streaming electron physics on the dynamics of magnetic reconnection. We also started to explore the applicability of anomalous‐resistivity‐type closures for magnetic reconnection in collisionless low‐β pair plasmas. In preliminary investigations of pair plasma and Reduced‐MHD (RMHD) island coalescence, we found that pair plasma simulations typically have very fast reconnection rates compared to RMHD simulations. We hypothesize that a closure 2 for the pair plasma system might take the form of an anomalous resistivity and have tested a simple form of this closure by increasing hyper‐dissipation coefficients globally. Although this recovers the kinetic reconnection rate in RMHD, it does not capture important island coalescence dynamics such as the generation of turbulence in the region surrounding the islands. In the future, machine learning could be used to discover more sophisticated anomalous resistivity models which can better capture these effects. Finally, with the goal of establishing better multi‐scale descriptions of collisionless plasmas, such as those associated with magnetic reconnection, we have also explored the use of machine learning to develop reduced models of particle injection. This is particularly important in plasmas where a large fraction of the energy is channeled into nonthermal particles. We have shown that deep neural networks can learn to parametrize injection by accurately describing the velocity moments of the injected particles in terms of a nonlinear combination of the moments of the background plasma. This enables the development of more accurate hybrid fluid‐kinetic models, where a fluid description is used to evolve the background plasma and a kinetic description is used for the nonthermal particles. Previous hybrid algorithms cannot capture the self‐consistent exchange of mass/charge between both descriptions. This problem can potentially be solved by this new method. 2.2 Graduate students research carried out on this project This grant supported an MIT graduate student who carried out research focused on numerical simulations of kink‐driven turbulence in a cylindrical pinch as a prelude to investigating reconnection driven by the coalescence of two such structures. The grant also supported the doctoral research of an MIT graduate student who studied the role of ion acoustic instability in magnetic reconnection [Li 2026]; the doctoral research of an MIT graduate student on electron- only magnetic reconnection and inverse magnetic-energy transfer at subion scales [Liu 2025]; and the research of an MIT post-doctoral associate on using quantized tensor networks for solving the Vlasov–Maxwell equations [Ye 2024]. These research activities are described in more detail in the following subsections. 2.2.1 Role of ion acoustic instability in magnetic reconnection A first-principles numerical study of magnetic reconnection in plasmas with different initial ion- to-electron temperature ratios was carried out [Li 2026]. In cases where this ratio is significantly below unity, intense wave activity was observed in the diffusion region, driven by the ion- acoustic instability (IAI). An example of this is shown in Fig. 1 where we have plotted time evolution of the temporally averaged reconnection rate for Ti0/Te0 =1/50 (blue), Ti0/Te0 =1/10 (orange) and Ti0/Te0 =1 (purple). AzX and AzO are the out-of-plane magnetic vector potentials at the X and O points, respectively. These simulations were carried out with the Osiros PIC code [Fonseca 2002] with initial conditions set to a Harris sheet equilibrium [Harris 1962], with a model initial magnetic vector potential. 3 Analysis showed that the dominant macroscopic effect of this instability is to drive substantial ion heating, raising temperatures to about one-tenth of electron temperature in the current sheet. Figure 1: Time evolution of the temporally averaged reconnection rate for Ti0/Te0 =1/50 (blue), Ti0/Te0 =1/10 (orange) and Ti0/Te0 =1 (purple). AzX and AzO are the out-of-plane magnetic vector potentials at the X and O points, respectively, and the semi-transparent curves are more exact values of the solid curves. Here tA is the global Alfven time. Reproduced from [Li 2026]. In contrast to earlier studies reporting significant anomalous resistivity, it was found that anomalous contributions to reconnection due to the IAI are minimal. Indeed, the observed ion heating was found to be driven mainly by reconnection electric fields, not turbulence-induced anomalous effects. Phenomenologically, the IAW’s develop into nonlinear ion phase-space holes, indicating saturation through ion scattering and electron plateau formation. The wave spectra are found to align with Kadomtsev–Petviashvili (KP) theory, confirming strong nonlinear effects in wave saturation. 2.2.2 Electron-only magnetic reconnection and inverse magnetic-energy transfer at subion scales In highly magnetized, low-β plasmas, such as those found in the Earth’s magnetosheath, magnetic reconnection may proceed in sufficiently small regime where the ions are effectively unresponsive, and the dynamics are dominated by electrons. An analytical model was derived 4 and numerically validated for “electron-only” magnetic reconnection, demonstrating that when the reconnection region is below the ion scales, the reconnection rate can significantly exceed the classical ion-coupled value [Liu 2025]. This is shown in Fig. 2, where the simulated values of peak reconnection rate from each simulation is plotted versus the normalized ion Larmor radius. Figure 2: Dimensionless peak reconnection rates, from simulations with different values of ion Larmor radius. In the electron-only regime, different color symbols refer to simulations with different electron skin depth. Reproduced from [Liu 2025]. This model was then applied to the phenomenon of inverse magnetic-energy transfer via island coalescence at sub-ion scales, obtaining new time-dependent scaling laws for the decay of magnetic energy and for the growth of magnetic structure sizes that differ from those established in magnetohydrodynamic (MHD) regimes. Using two- and three-dimensional kinetic simulations, the authors showed that magnetic structures born at sub-ion scales can coalesce and grow until they reach system-size dimensions. This result demonstrates a reconnection-mediated route to large-scale magnetic structure formation which lies outside the standard MHD paradigm. 2.2.3 Quantized tensor networks for solving the Vlasov–Maxwell equations The Vlasov–Maxwell equations provide an ab initio description of collisionless plasmas but solving them is often impractical because of the wide range of spatial and temporal scales that must be resolved and the high dimensionality of the problem. In this research, a quantum- inspired semi-implicit Vlasov–Maxwell solver was developed [Ye 2024] that uses the quantized tensor network (QTN) framework. With this QTN solver, the cost of grid-based numerical simulation of size N is reduced from O(N) to O(poly(D)), where D is the ‘rank’ or ‘bond dimension’ of the QTN and is typically set to be much smaller than N. This method was tested on five-dimensional plasma problems, including the Orszag–Tang vortex and GEM magnetic reconnection problem. It was found that for the five-dimensional (2-space, 3- velocity) test problems considered, a modest D = 64 appears to be sufficient for capturing the expected physics despite the simulations using a total of N = 236 grid points, which would require 5 D = 218 for full-rank calculations. Additionally, it was observed that a QTN time evolution scheme based on the Dirac–Frenkel variational principle allows one to use somewhat larger time steps than prescribed by the Courant–Friedrichs–Lewy constraint. Figure 3: Out-of-plane current density for the GEM reconnection problem. Calculations were performed with bond dimension (a) D = 32 and (b) D = 64, and time step t = 0.005Ω−1c,p. Time is in units of Ω−1c,p. Reproduced from [Ye 2024]. Figure 3 shows the current density calculations for the GEM reconnection problem with bond dimensions D = 32 and D = 64. The simulation yields roughly the expected reconnected flux over time [see Schmitz & Grauer 2006]. However, the dynamics appear sensitive to numerical noise, as evidenced by unexpected secondary island formation in the Harris sheet for the D = 32 case. In summary, the QTN format is a promising means of approximately solving the Vlasov– Maxwell equations with significantly reduced cost. This method captures key plasma physics phenomena with significantly fewer degrees of freedom, thus offering the potential for large- scale, high-dimensional plasma simulations and generating training data for machine learning. 6 References [Alves & Fiuza 2022] E. P. Alves and F. Fiuza, “Data-driven discovery of reduced plasma physics models from fully kinetic simulations”, Physical Review Research 4, 033192 (2022). [Fonseca 2002] R. A. Fonseca et al, “Osiris: A three-dimensional, fully relativistic particle in cell code for modeling plasma based accelerators. Proceedings of the International Conference on Computational Science 2331, 342–351 (2002). [Harris 1962] E. G. Harris, “On a plasma sheath separating regions of oppositely directed magnetic field”, Il Nuovo Cimento 23, 115–121, 1955 (1962). [Li 2026] D. Li, Z. Liu, and N. F. Loureiro, “Role of ion acoustic instability in magnetic reconnection”, Journal of Plasma Physics 92, E20 (2026); https://doi.org/10.1017/S002237782510113X [Liu 2025] Z. Liu, C. Silva, L. M. Milanese, M. Zhou, N. R. Mandell, and N. F. Loureiro, “Electron-Only Magnetic Reconnection and Inverse Magnetic-Energy Transfer at Subion Scales”, Physical Review Letters 134, 155201 (2025); DOI: 10.1103/PhysRevLett.134.155201 [Schmitz & Grauer 2006] H. Schmitz & R. Grauer, “Kinetic Vlasov simulations of collisionless magnetic reconnection”, Physics of Plasmas 13 (9), 092309 (2006). [Ye 2024] E. Ye and N. F. Loureiro, “Quantized tensor networks for solving the Vlasov–Maxwell equations”, Journal of Plasma Physics 90, 805900301 (2024); doi:10.1017/S0022377824000503 7