ADITYA-U Tokamak Overview

• ADITYA-U Tokamak is a medium-sized air-core, ohmically heated device characterized by versatile limiter and diverted equilibria and rapid short-pulse operation.
• The device integrates real-time gas fueling, auxiliary RF systems, and advanced diagnostics to modify plasma transport and stabilize MHD phenomena.
• Research on ADITYA-U employs machine learning and equilibrium reconstruction to predict disruptions and optimize plasma performance with high accuracy.

ADITYA-U Tokamak is an upgraded, medium-sized tokamak at the Institute for Plasma Research in Gandhinagar, India, described across recent studies as an air-core, ohmically heated device with major radius $R = 0.75~\mathrm{m}$, minor radius $a = 0.25~\mathrm{m}$, and aspect ratio $R/a = 3$. Reported operating ranges include plasma current $I_p = 80\text{–}200~\mathrm{kA}$, toroidal magnetic field $B_t = 0.9\text{–}1.2~\mathrm{T}$, central chord-averaged density $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$, and core electron temperature $200\text{–}400~\mathrm{eV}$, with pulse duration about $300~\mathrm{ms}$ in negative converter operation (Banerjee et al., 23 Jul 2025). Recent work on ADITYA-U spans plasma start-up control, auxiliary-wave coupling, edge and core transport, sawtooth and disruption physics, equilibrium reconstruction, and the development of new diagnostics and data-driven surrogates, indicating a research program that couples machine-specific experiments with modeling and real-time control methods (Maurya et al., 6 Jul 2026).

1. Device configuration and operational regime

ADITYA-U is reported as operating in both limiter and diverted equilibria, depending on the study and the physics problem being addressed. Limiter operation appears prominently in studies of edge turbulence, scrape-off-layer transport, and reciprocating-probe measurements, with a toroidal belt limiter on the high-field side and two poloidal limiters on the low-field side in the circular limiter configuration modeled with UEDGE (Dey et al., 27 Mar 2025). A diverted equilibrium is explicitly noted in the gyrokinetic study of transient fueling and turbulence suppression (Alageshan et al., 23 Feb 2026).

The machine is described as being transformer driven for plasma breakdown and Ohmic current drive, with typical breakdown loop voltage about $18\text{–}20~\mathrm{V}$ and flat-top loop voltage about $1.8\text{–}2~\mathrm{V}$. Typical Ohmic operation has edge safety factor $a = 0.25~\mathrm{m}$0, while maximum achieved plasma current around $a = 0.25~\mathrm{m}$1 can reduce $a = 0.25~\mathrm{m}$2 to about $a = 0.25~\mathrm{m}$3 (Aich et al., 12 Jan 2026). In separate disruption studies, the discharges examined had edge safety factor $a = 0.25~\mathrm{m}$4, allowing the $a = 0.25~\mathrm{m}$5 rational surface to lie near the plasma edge in some cases and deeper inside in others (Banerjee et al., 23 Jul 2025).

Several papers emphasize the consequences of the machine’s short-pulse character. One disruption-prediction study states that ADITYA has average shot duration around $a = 0.25~\mathrm{m}$6, while a later transformer-based study notes about $a = 0.25~\mathrm{m}$7 for ADITYA and about $a = 0.25~\mathrm{m}$8 for ADITYA-U, making millisecond-scale prediction and actuation windows operationally important (Agarwal et al., 17 Jul 2025). This suggests that ADITYA-U occupies an intermediate scale in which control, diagnostics, and inference must be fast enough for online use, but are still constrained by relatively short discharges and limited available diagnostics.

2. Equilibrium formation, current ramp-up, and auxiliary-wave operation

A central operational issue in ADITYA-U is plasma current ramp-up. A recent control study states that start-up proceeds through breakdown, burn-through, ramp-up, and flat-top, and identifies the ramp-up phase as especially sensitive because the plasma column is still evolving rapidly and is easy to destabilize. The paper further reports that the primary vertical-field coil time constant is about $a = 0.25~\mathrm{m}$9, whereas plasma column motion evolves on a time scale of about $R/a = 3$0, so equilibrium control can become inadequate when $R/a = 3$1 becomes too large (Dolui et al., 4 Jul 2026).

To address this, a real-time active control scheme using neutral gas injection was developed. In that scheme, a dedicated DSP-based controller hardware monitors the Rogowski-coil-derived plasma current slope every $R/a = 3$2. When the rise rate exceeds a threshold, the controller generates a TTL pulse that triggers a function generator and then a piezoelectric valve at the bottom port, producing a short hydrogen gas puff. The calibration relation given is

$R/a = 3$3

so that

$R/a = 3$4

Using the $R/a = 3$5 slope window, a $R/a = 3$6 change corresponds to about

$R/a = 3$7

In the plasma experiment, when the current rise rate exceeded about $R/a = 3$8 and reached around $R/a = 3$9, a $I_p = 80\text{–}200~\mathrm{kA}$0 gas puff was triggered, reducing the rise rate to below $I_p = 80\text{–}200~\mathrm{kA}$1. A direct comparison between discharge #38970 with gas injection active and discharge #38971 with gas injection inactive showed closer agreement between required and preset vertical fields, reduced $I_p = 80\text{–}200~\mathrm{kA}$2 spikes and MHD activity, and a more desirable peaked SXR profile with sawtooth oscillations in the controlled case (Dolui et al., 4 Jul 2026).

ADITYA-U is also equipped with auxiliary RF systems. One paper reports a $I_p = 80\text{–}200~\mathrm{kA}$3, $I_p = 80\text{–}200~\mathrm{kA}$4 Electron Cyclotron Resonant Heating system launched from the low-field side in O-mode from port 14, and a $I_p = 80\text{–}200~\mathrm{kA}$5, $I_p = 80\text{–}200~\mathrm{kA}$6 Lower Hybrid Current Drive system launched from the low-field side using a Passive Active Multijunction launcher from port 5 (Aich et al., 12 Jan 2026). Successful ECRH coupling is reported to increase electron heating, diamagnetic current, $I_p = 80\text{–}200~\mathrm{kA}$7, and stored energy, while pushing the plasma column toward the low-field side. Effective coupling is stated to occur above a critical $I_p = 80\text{–}200~\mathrm{kA}$8 of about $I_p = 80\text{–}200~\mathrm{kA}$9 and to be preferred around $B_t = 0.9\text{–}1.2~\mathrm{T}$0. LHCD, by contrast, is reported to have little effect on diamagnetism in the measured operational range, but to broaden the current profile and reduce internal inductance $B_t = 0.9\text{–}1.2~\mathrm{T}$1 when coupling is good; coupling is best when the plasma is shifted outward toward the low-field side (Aich et al., 12 Jan 2026).

These studies jointly depict ADITYA-U as a device in which equilibrium maintenance is tightly linked to dynamic current evolution, gas fueling, and actuator placement. A plausible implication is that the machine serves as a platform for testing integrated control strategies in which fueling and RF systems act not only as particle and power sources but also as equilibrium and current-profile actuators.

3. Fueling, turbulence, and transport modification

A recurring theme in ADITYA-U research is the use of short gas puffs as active transport-control actuators rather than merely as fueling sources. In flat-top discharges, a piezo valve is reported to inject about $B_t = 0.9\text{–}1.2~\mathrm{T}$2 molecules for a $B_t = 0.9\text{–}1.2~\mathrm{T}$3 voltage pulse, and the resulting perturbation produces a characteristic sequence: electron density rises rapidly and peaks in about $B_t = 0.9\text{–}1.2~\mathrm{T}$4, whereas core electron temperature rises later, peaking about $B_t = 0.9\text{–}1.2~\mathrm{T}$5 after injection (Alageshan et al., 23 Feb 2026).

The principal profile effect is not a uniform density increase but a flattening of the radial density profile near mid-radius. In the sawtooth-control study, the density profile becomes flatter mainly because density increases in the mid-radius region while the core density changes little (Dolui et al., 3 Jan 2025). In the gyrokinetic transport study, the affected region is stated more explicitly as $B_t = 0.9\text{–}1.2~\mathrm{T}$6, with the core remaining nearly constant for $B_t = 0.9\text{–}1.2~\mathrm{T}$7. Global electrostatic gyrokinetic simulations with the GTC code, using equilibria from IPREQ and experimental profiles, identify trapped electron mode turbulence as the dominant instability. The reported linear growth rate changes only slightly, from $B_t = 0.9\text{–}1.2~\mathrm{T}$8 before the puff to $B_t = 0.9\text{–}1.2~\mathrm{T}$9 after it, but the peak of the flux-surface-averaged mode RMS shifts outward from $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$0 to $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$1, and the mode number decreases from about $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$2 to about $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$3. The nonlinear consequences are substantial: the saturation level in electron diffusivity $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$4 is reduced by about $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$5, the saturation level in ion diffusivity $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$6 is reduced by about $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$7, and $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$8 remains about an order of magnitude larger than $(1\text{–}4)\times 10^{19}~\mathrm{m}^{-3}$9, consistent with TEM-dominated transport (Alageshan et al., 23 Feb 2026).

The same physical chain is used to explain sawtooth stabilization by short gas pulses. In Ohmically heated hydrogen discharges with $200\text{–}400~\mathrm{eV}$0, $200\text{–}400~\mathrm{eV}$1, and $200\text{–}400~\mathrm{eV}$2, repeated gas puffs at about $200\text{–}400~\mathrm{eV}$3 intervals were found to increase the sawtooth ramp phase from about $200\text{–}400~\mathrm{eV}$4 to about $200\text{–}400~\mathrm{eV}$5, roughly a factor of 2 increase. The inversion radius was reported as $200\text{–}400~\mathrm{eV}$6, corresponding to $200\text{–}400~\mathrm{eV}$7, and sawtooth crashes were found to occur when

$200\text{–}400~\mathrm{eV}$8

near that radius, irrespective of whether gas puffing was applied. The interpretation offered is that gas puffing suppresses TEM-driven turbulence, reduces heat diffusivity, slows the post-crash recovery of the temperature gradient at $200\text{–}400~\mathrm{eV}$9, and thereby delays the next crash (Dolui et al., 3 Jan 2025).

Fueling-induced transport modification also appears in the edge and SOL. A pulsed-fueling study reports that short periodic hydrogen pulses of $300~\mathrm{ms}$0 molecules broaden the scrape-off-layer heat-flux width by a factor of $300~\mathrm{ms}$1, while reducing edge heat flux near the LCFS from about $300~\mathrm{ms}$2 in continuous density-control discharges to about $300~\mathrm{ms}$3 in the pulse-driven case (Hoque et al., 1 Aug 2025). The paper writes

$300~\mathrm{ms}$4

and argues analytically that temperature reduction alone cannot explain the observed broadening. In the analytic model, holding $300~\mathrm{ms}$5 fixed and reducing $300~\mathrm{ms}$6 from $300~\mathrm{ms}$7 to $300~\mathrm{ms}$8 fails to reproduce the measured profile, whereas increasing $300~\mathrm{ms}$9 from $18\text{–}20~\mathrm{V}$0 to $18\text{–}20~\mathrm{V}$1 gives much better agreement. UEDGE simulations further indicate that the best post-pulse agreement requires $18\text{–}20~\mathrm{V}$2 together with a stronger inward pinch velocity $18\text{–}20~\mathrm{V}$3 (Hoque et al., 1 Aug 2025).

The inward-pinch picture is consistent with a separate UEDGE study of circular limiter plasmas, which concludes that the measured edge density profile can only be reproduced with a constant inward convective velocity

$18\text{–}20~\mathrm{V}$4

in combination with

$18\text{–}20~\mathrm{V}$5

That work explicitly states that diffusion alone is insufficient and relates the fitted inward convection to a Ware-pinch estimate of order $18\text{–}20~\mathrm{V}$6 (Dey et al., 27 Mar 2025). Taken together, these results suggest a device-level picture in which short gas puffs reshape both core and edge transport by profile flattening, enhanced cross-field diffusion, and inward convection, with beneficial effects on confinement and heat-flux distribution.

4. MHD activity, sawteeth, and disruption phenomenology

MHD dynamics in ADITYA-U span edge-coupled coherent modes, sawtooth oscillations, magnetic islands, and multiple disruption classes. In limiter discharges, one study reports that when the amplitude of the dominant $18\text{–}20~\mathrm{V}$7 resistive MHD mode exceeds about $18\text{–}20~\mathrm{V}$8, coherent oscillations appear in edge plasma potential and density fluctuations at the same frequency as the MHD mode. Multi-point Langmuir-probe analysis yields $18\text{–}20~\mathrm{V}$9 for the potential fluctuation and $1.8\text{–}2~\mathrm{V}$0 for the density fluctuation, with both structures having $1.8\text{–}2~\mathrm{V}$1. The coherent frequencies are reported in the $1.8\text{–}2~\mathrm{V}$2 range, while the dominant MHD activity itself lies at about $1.8\text{–}2~\mathrm{V}$3. The paper interprets the disparate mode structures as consistent with excitation of a global high-frequency branch of zonal flows or a GAM-like mode through coupling of even harmonics of potential to odd harmonics of pressure due to the $1.8\text{–}2~\mathrm{V}$4 dependence of the toroidal magnetic field (Singh et al., 2024).

Sawtooth dynamics are treated separately but are closely connected to the same internal transport landscape. The sawtooth-control study reports quiescent ramp phase $1.8\text{–}2~\mathrm{V}$5 and crash phase $1.8\text{–}2~\mathrm{V}$6 without gas puffing, with $1.8\text{–}2~\mathrm{V}$7-like odd-parity precursor oscillations showing $1.8\text{–}2~\mathrm{V}$8 phase difference across the core. After short gas-pulse injection, the ramp phase extends to $1.8\text{–}2~\mathrm{V}$9, and the SXR intensity at the inversion region returns to pre-crash levels in about $a = 0.25~\mathrm{m}$00 with gas puff, compared with about $a = 0.25~\mathrm{m}$01 without gas puff (Dolui et al., 3 Jan 2025). The new thin-target hard X-ray diagnostic later developed for Aditya-U was designed specifically to examine fast electrons generated during sawtooth activity and to distinguish core-confined thin-target bremsstrahlung from limiter- or wall-associated thick-target emission (Dolui et al., 1 Jul 2026).

Disruption research has identified a new regime distinct from the conventional locked-mode disruption class. A statistical study of 150 disruptive discharges introduces Accelerated Mode Disruption as a regime governed predominantly by the $a = 0.25~\mathrm{m}$02 drift-tearing mode. In AMD, the mode frequency rises monotonically while the amplitude saturates nonlinearly, followed by a sudden frequency collapse and sharp amplitude increase. In shot #37103, the Mirnov frequency is reported to rise by about $a = 0.25~\mathrm{m}$03 over the precursor window, then fall from about $a = 0.25~\mathrm{m}$04 to $a = 0.25~\mathrm{m}$05 in $a = 0.25~\mathrm{m}$06, coincident with island growth from about $a = 0.25~\mathrm{m}$07 to $a = 0.25~\mathrm{m}$08. Statistical separation from LMD is based on empirical thresholds: $a = 0.25~\mathrm{m}$09 of AMDs occur when $a = 0.25~\mathrm{m}$10, whereas $a = 0.25~\mathrm{m}$11 of LMDs occur when $a = 0.25~\mathrm{m}$12; AMD is associated with current drop more than $a = 0.25~\mathrm{m}$13, current quench time $a = 0.25~\mathrm{m}$14, and CQ rate $a = 0.25~\mathrm{m}$15 (Banerjee et al., 23 Jul 2025).

The same study attributes AMD to core radiation increase, core temperature hollowing, and steepening of both pressure gradient and current density profile near the rational surface. The diamagnetic contribution to the DTM frequency is quantified as dominant, about $a = 0.25~\mathrm{m}$16 of the total for shot #37103, with toroidal flow contributing about $a = 0.25~\mathrm{m}$17 and poloidal flow about $a = 0.25~\mathrm{m}$18. This suggests that disruption precursors in ADITYA-U are not reducible to simple wall-locking scenarios and may depend sensitively on evolving internal profiles (Banerjee et al., 23 Jul 2025).

Magnetic-island effects also extend into microinstability physics. A gyrokinetic study using G2C3 examines static $a = 0.25~\mathrm{m}$19 and $a = 0.25~\mathrm{m}$20 islands relevant to ADITYA-U, with experimentally observed $a = 0.25~\mathrm{m}$21 island width about $a = 0.25~\mathrm{m}$22. In the first simulation phase, the island topology causes density flattening inside the island region; in the second, the relaxed profiles are used for linear electrostatic ITG calculations with adiabatic electrons. For the $a = 0.25~\mathrm{m}$23 case, density flattening becomes visible for island width $a = 0.25~\mathrm{m}$24. As island width grows, growth rates converge to

$a = 0.25~\mathrm{m}$25

and the fluctuation structure is suppressed in the island core and enhanced near the separatrix. The study concludes that islands stabilize ITG through profile flattening and geometric restructuring, with the $a = 0.25~\mathrm{m}$26 island producing a broader mode structure and a slightly smaller linear growth rate than the $a = 0.25~\mathrm{m}$27 case (Singh et al., 3 Feb 2026).

5. Diagnostic systems and measurement infrastructure

ADITYA-U has recently seen a marked expansion of diagnostic capability. A notable development is a thin-target hard X-ray bremsstrahlung detection system designed specifically to resolve the ambiguity between confined and lost runaway-electron signals during sawtooth events. The pre-existing monitor was a NaI(Tl) scintillation detector about $a = 0.25~\mathrm{m}$28 from the tokamak, viewing tangentially across essentially the entire vacuum vessel and therefore collecting a mixture of thin-target plasma bremsstrahlung and thick-target wall or limiter emission. The new system instead uses a specially shielded CdTe detector with a lead collimator and additional filtering, with a restricted line of sight to the plasma core including the sawtooth inversion radius (Dolui et al., 1 Jul 2026).

The detector is described as a $a = 0.25~\mathrm{m}$29 thick CdTe crystal with active area $a = 0.25~\mathrm{m}$30, effective over roughly $a = 0.25~\mathrm{m}$31, housed in $a = 0.25~\mathrm{m}$32 of aluminum filtering and enclosed in a $a = 0.25~\mathrm{m}$33 thick, $a = 0.25~\mathrm{m}$34 long lead shield. The lead collimator has an $a = 0.25~\mathrm{m}$35 aperture, $a = 0.25~\mathrm{m}$36 length, and $a = 0.25~\mathrm{m}$37 outer diameter, and defines a field of view of about $a = 0.25~\mathrm{m}$38. A key validation experiment compared the signal with the collimator aperture open and blocked by a lead rod; when blocked, the HXR signal nearly vanished, showing that measured counts in the open case were dominated by photons entering through the intended line of sight. Forward modeling then compared synthetic and measured pulse-height spectra, with best agreement obtained for a runaway-electron beam energy of about $a = 0.25~\mathrm{m}$39 and pitch angle near $a = 0.25~\mathrm{m}$40 (Dolui et al., 1 Jul 2026).

Edge diagnosis has likewise been strengthened by the development of a high-speed reciprocating drive system with interchangeable Langmuir and magnetic probe heads. The HRDS is a servo-motor-driven system installed through a $a = 0.25~\mathrm{m}$41 CF gate valve, with a $a = 0.25~\mathrm{m}$42 stainless-steel edge-welded bellows, $a = 0.25~\mathrm{m}$43 stroke length, and remote shot-to-shot control of speed, acceleration, scan length, and timing. Required scan distance during a discharge was $a = 0.25~\mathrm{m}$44 in $a = 0.25~\mathrm{m}$45, corresponding to required average speed $a = 0.25~\mathrm{m}$46; achieved average operating speed was about $a = 0.25~\mathrm{m}$47, with achieved acceleration about $a = 0.25~\mathrm{m}$48. The system was tested to a leak rate of about $a = 0.25~\mathrm{m}$49, with electrical isolation resistance $a = 0.25~\mathrm{m}$50 at $a = 0.25~\mathrm{m}$51 (Singh et al., 8 Jan 2025).

Using an 8-tip rake Langmuir probe and then a 5-probe magnetic head, the HRDS extended edge measurements from the previous limit of about $a = 0.25~\mathrm{m}$52 inside the limiter radius to penetration about $a = 0.25~\mathrm{m}$53, reaching about $a = 0.25~\mathrm{m}$54. In tokamak-plasma operation, motion of the probe head caused plasma current variation $a = 0.25~\mathrm{m}$55 and no noticeable effect on SXR, edge density, or edge potential during the current flat-top. Measurements showed that a fuel puff of about $a = 0.25~\mathrm{m}$56 molecules of $a = 0.25~\mathrm{m}$57 affects plasma parameters up to about $a = 0.25~\mathrm{m}$58 inside the LCFS, flattening radial density and temperature profiles and suppressing fluctuations. The magnetic probe head additionally provided poloidal magnetic fluctuation measurements inside the LCFS that agreed with a fixed Mirnov coil at $a = 0.25~\mathrm{m}$59 (Singh et al., 8 Jan 2025).

On the data-management side, PRISM, a MATLAB-based application developed in App Designer and tested with ADITYA-U data, provides structured registration, retrieval, and 2D/3D visualization of probe metadata. It uses Microsoft Excel .xlsx files on a shared network drive, with one worksheet per probe type, and supports filtering by date, shot number, position, shot duration, or experiment. The application captures fields such as probe name, date, shot number, shot duration, port type, port number, radial position $a = 0.25~\mathrm{m}$60, $a = 0.25~\mathrm{m}$61, $a = 0.25~\mathrm{m}$62, channel number, information type, experiment, and remarks (Verma et al., 30 Aug 2025). This suggests that diagnostic sophistication in ADITYA-U is increasingly accompanied by formalized metadata infrastructure.

6. Equilibrium reconstruction, machine-learning surrogates, and disruption prediction

Equilibrium analysis in ADITYA-U has advanced through both conventional Grad–Shafranov reconstruction and learned surrogates. The Python free-boundary equilibrium code pyIPREQ solves

$a = 0.25~\mathrm{m}$63

using finite differences, Green’s functions, and Picard iteration, while supporting limiter boundaries and equilibria constrained by prescribed magnetic-axis position. For ADITYA-U, this is particularly relevant because magnetic-axis location is available from Sine-Cosine diagnostics. In a benchmark against the original IPREQ for a hypothetical ADITYA-U case with $a = 0.25~\mathrm{m}$64, $a = 0.25~\mathrm{m}$65, and $a = 0.25~\mathrm{m}$66, pyIPREQ reproduced the magnetic axis to high accuracy: $a = 0.25~\mathrm{m}$67 in IPREQ versus $a = 0.25~\mathrm{m}$68 in pyIPREQ, with $a = 0.25~\mathrm{m}$69 versus $a = 0.25~\mathrm{m}$70 (Maurya et al., 24 Jul 2025).

A later study builds deep-learning surrogates on top of such equilibrium calculations. Using pyIPREQ, a synthetic free-boundary equilibrium dataset of 100,760 cases was generated from 766 ADITYA-U discharges spanning 2021–2025, restricted to circular limiter plasmas near flat-top and defined on a rectangular $a = 0.25~\mathrm{m}$71-$a = 0.25~\mathrm{m}$72 grid with $a = 0.25~\mathrm{m}$73, $a = 0.25~\mathrm{m}$74, and spatial resolution $a = 0.25~\mathrm{m}$75. The models predict scalar quantities, 1D safety-factor profiles, 2D poloidal flux profiles, and inverse coil-current settings (Maurya et al., 6 Jul 2026).

For magnetic-axis prediction, a Dense network using 26 inputs achieved median absolute test errors of about $a = 0.25~\mathrm{m}$76 for $a = 0.25~\mathrm{m}$77 and about $a = 0.25~\mathrm{m}$78 for $a = 0.25~\mathrm{m}$79. Separate scalar models predicted $a = 0.25~\mathrm{m}$80 with median error about $a = 0.25~\mathrm{m}$81, $a = 0.25~\mathrm{m}$82 with median error about $a = 0.25~\mathrm{m}$83, and $a = 0.25~\mathrm{m}$84 with median error about $a = 0.25~\mathrm{m}$85. For $a = 0.25~\mathrm{m}$86, a PCA reduced-order model used 4 PCA modes to capture $a = 0.25~\mathrm{m}$87 of the variance, while a 1D CNN predicted $a = 0.25~\mathrm{m}$88 and reconstructed $a = 0.25~\mathrm{m}$89 by integration from the edge inward. For $a = 0.25~\mathrm{m}$90, both PCA-based and 2D CNN models incorporated Grad–Shafranov residual constraints through

$a = 0.25~\mathrm{m}$91

with the largest 2D CNN model taking about $a = 0.25~\mathrm{m}$92 per equilibrium instance on CPU (Maurya et al., 6 Jul 2026). The paper explicitly frames these models as computationally efficient alternatives for real-time plasma control, rapid equilibrium analysis, and experimental planning.

Data-driven disruption prediction constitutes a parallel line of work. An earlier ADITYA study used a stateful LSTM on 119 disruptive shots and selected diagnostics including plasma current, loop voltage, bolometer probe, Mirnov, HXR, and SXR. After downsampling to a common interval of $a = 0.25~\mathrm{m}$93, the model predicted disruption $a = 0.25~\mathrm{m}$94 in advance, with training accuracy $a = 0.25~\mathrm{m}$95 and test accuracy $a = 0.25~\mathrm{m}$96. The test set of 36 shots yielded false alarm $a = 0.25~\mathrm{m}$97, premature alarm $a = 0.25~\mathrm{m}$98, missed alarm $a = 0.25~\mathrm{m}$99, and true alarm $R/a = 3$00, with inference time under $R/a = 3$01 per time step on an Intel Xeon processor (Agarwal et al., 2020).

A later transformer-based study, motivated explicitly by the short-pulse regime of ADITYA and ADITYA-U, used six diagnostic signals—plasma current, SXR, HXR, bolometer, $R/a = 3$02, and $R/a = 3$03-III—resampled to $R/a = 3$04, on a dataset filtered to 415 disruptive and 310 non-disruptive shots from 1407 experiments. The transformer encoder outperformed the LSTM baseline in recall and accuracy across lead times from $R/a = 3$05 to $R/a = 3$06, maintaining recall above $R/a = 3$07 up to about $R/a = 3$08, and remained robust up to an $R/a = 3$09 lead time (Agarwal et al., 17 Jul 2025). This suggests that ADITYA-U has become a testbed not only for plasma control itself but also for real-time inference architectures adapted to short-duration tokamak discharges.