Tatva: Elemental Principles & FEM Framework

• Tatva is defined as both a classical notion of elemental principles in Indian thought and a modern computational framework using a single energy functional.
• The historical aspect organizes reality via Panchmahabhootas, reflecting early efforts to explain diversity by reducing it to fundamental constituents.
• The computational tatva framework employs global automatic differentiation and batched operator strategies to achieve scalable, efficient GPU-based finite-element simulations.

Tatva, also written Tattva in the historical-literary context, denotes “the elemental principles that constitute reality” in classical Indian thought, and in contemporary computational mechanics it is also the name of an energy-centric finite-element framework. In the former sense, Tatva appears as part of the Panchmahabhootas tradition of elemental categories; in the latter, tatva represents physics through a single global energy functional and applies automatic differentiation globally to obtain residual and tangent operators for large-scale GPU-native finite-element computation (Godbole, 2010, Pundir et al., 12 Feb 2026). The shared lexical root is continuity of emphasis on underlying principles rather than continuity of method: the historical notion is qualitative and metaphysical, whereas the computational framework is variational, algorithmic, and explicitly implemented in JAX/XLA.

1. Terminological scope and conceptual split

In "The Heart of Matter" (Godbole, 2010), Tatva (Tattva) is introduced as a classical notion of elemental principles. The paper places it at the beginning of a long intellectual arc extending from Panchmahabhootas and Empedocles’ elements to quarks, leptons, and gauge bosons. Its role there is historical and conceptual: it names an early conviction that the diversity of observed phenomena rests on a smaller set of underlying principles.

In "A versatile FEM framework with native GPU scalability via globally-applied AD" (Pundir et al., 12 Feb 2026), tatva is a software and methodological framework for computational mechanics. It is defined as an energy-centric finite-element framework in which the physics of a problem is represented entirely by a single global scalar functional $E(u)$, and automatic differentiation is applied to that global functional to generate residual and tangent operators. The framework is implemented in JAX with GPU-native execution via XLA.

These two usages are semantically related only at the level of etymological resonance. A plausible implication is that the framework name deliberately evokes a “first-principles” style of modeling, but the technical content of the two papers is otherwise distinct.

2. Tatva in classical Indian thought and the historical search for constituents

In the historical account of particle physics given in (Godbole, 2010), Tatva/Tattva refers to elemental principles, specifically in connection with the Panchmahabhootas: earth, water, fire, air, and sky/ether. The paper juxtaposes these categories with Empedocles’ four elements and then contrasts both with the Standard Model’s empirically grounded catalog of quarks, leptons, and gauge bosons. The continuity identified in that account is not the content of the elemental lists, but the recurring attempt to determine the “bricks and mortar” of nature and the forces that bind them.

The paper emphasizes that Panchmahabhootas were qualitative categories used to organize manifest phenomena. They expressed the conviction that there are elemental principles behind the world’s diversity, but did not provide an empirical method for isolating those principles. By contrast, the subsequent scientific trajectory is organized around quantitative regularities, falsifiable hypotheses, and scattering experiments as probes of structure.

This distinction is explicit in the paper’s treatment of convergence and divergence between Tatva-like elemental thought and modern particle physics. Both seek a compact set of fundamentals from which diversity emerges, but particle physics separates constituents from force carriers, requires a mathematical dynamical framework, and treats “elementarity” operationally: what remains point-like and universal under the best available probes. This suggests that the meaning of Tatva in physics is not fixed by metaphysical category, but by experimental resolution.

3. From Tatva to the Standard Model: empirical elementarity

The historical narrative in (Godbole, 2010) presents the transition from philosophical elementality to empirical particle physics through a sequence of increasingly precise probes. Faraday’s work on discrete charge, J. J. Thomson’s isolation of the electron, Rutherford’s beam–target–detector gold-foil scattering experiments, Hofstadter’s electron scattering measurements, and deep inelastic scattering collectively recast the question of “what matter is made of” into one of measurable substructure.

A central relation in that account is the de Broglie wavelength,

$\lambda = \frac{h}{2\pi p},$

with $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$. The paper uses this to explain why higher-momentum probes resolve finer scales: substructure is resolved only when the probe wavelength is comparable to or smaller than the target size. Rutherford’s MeV $\alpha$ particles resolved the atom to a compact nucleus; Hofstadter’s 400–600 MeV electrons established that nuclei have finite size roughly $10^4$–$10^5$ times smaller than atoms; and deep inelastic scattering at 10–20 GeV revealed point-like constituents inside the proton.

The paper then connects the hadron “zoo” to the quark idea. Gell-Mann and Zweig proposed fractionally charged quarks, and the refinement by Greenberg, Han, and Nambu introduced three colours per quark flavor. The specific example of the $\Delta^{++}$ is used to motivate colour: with electric charge $2e$ and spin $3/2$ in units of $\hbar$, it requires three $\lambda = \frac{h}{2\pi p},$0 quarks with aligned spins, and antisymmetry is restored only if those quarks carry distinct colour labels.

Deep inelastic scattering is described as decisive evidence for quarks and gluons. In elastic electron–proton scattering, a fixed outgoing electron energy is expected at a given angle; in DIS, a continuum of outgoing electron energies is observed,

$\lambda = \frac{h}{2\pi p},$1

because the electron scatters from moving constituents carrying fractions of the proton momentum. The paper states that later electron, muon, and neutrino probes confirmed quarks and gluons as point-like down to distances below about a thousandth of a fermi, while leptons showed no signs of substructure (Godbole, 2010).

4. Tatva as an energy-centric finite-element framework

In computational mechanics, tatva is defined by a “one energy, global AD” design philosophy (Pundir et al., 12 Feb 2026). The state of the system is described by a single scalar functional $\lambda = \frac{h}{2\pi p},$2, and all physics—materials, loads, constraints, and couplings—is written into that one energy. The governing equations are then obtained as the first and second variations of $\lambda = \frac{h}{2\pi p},$3 with respect to all degrees of freedom:

$\lambda = \frac{h}{2\pi p},$4

For continuous mechanical problems, the framework uses a global functional of the form

$\lambda = \frac{h}{2\pi p},$5

where $\lambda = \frac{h}{2\pi p},$6 is the bulk energy density, $\lambda = \frac{h}{2\pi p},$7 encodes boundary work or energetic boundary terms, and $\lambda = \frac{h}{2\pi p},$8 collects global terms such as penalties, constraints, or learned components. In discretized form, tatva sums scalar quadrature contributions over elements and quadrature points:

$\lambda = \frac{h}{2\pi p},$9

Surface and interface terms are handled similarly.

The paper positions this approach against locally applied AD and traditional assembly. Local differentiation of element-level residuals and tangents followed by scatter-add into global sparse operators scales well on CPUs, but the paper identifies scatter-add on GPUs as suffering from irregular memory access and atomic contention. Tatva instead differentiates the entire global energy functional. According to the paper, this preserves a direct correspondence to the variational formulation and makes it natural to encode nonlocal terms, mixed-dimensional couplings, multiphysics couplings, and constraints directly in $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$0.

The framework is implemented in JAX, uses JAX arrays and AD primitives such as jax.jacrev, jax.jvp, and jax.jacfwd, and relies on GPU compilation via XLA. Batched execution is implemented with jax.lax.map, specifically to ensure that peak memory depends on batch size rather than on total mesh size (Pundir et al., 12 Feb 2026).

5. Global automatic differentiation, JVPs, and sparse coloring

The principal technical claim of tatva is that globally applied AD can be made scalable on GPUs by combining batched evaluation of the global scalar functional with two complementary operator strategies: matrix-free Jacobian–vector products and coloring-based sparse differentiation (Pundir et al., 12 Feb 2026).

For matrix-free access, the tangent action on a vector is computed without materializing the tangent matrix:

$h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$1

Operationally, tatva applies reverse-mode AD to the global scalar $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$2 to obtain $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$3, then applies forward-mode AD to $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$4 to evaluate JVPs. The paper states that each JVP costs about the same as a residual evaluation and scales linearly with problem size. This enables Newton–Krylov linearization, including CG for SPD systems and GMRES otherwise, using only the operator action.

When an explicit sparse tangent matrix is required, tatva reconstructs it through distance-2 graph coloring of the mesh connectivity. The central idea is to compute multiple independent columns of $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$5 simultaneously by seeding a perturbation vector with canonical basis entries corresponding to a color class. If $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$6 is the number of colors, the paper gives the approximate cost relation

$h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$7

with $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$8 depending on local connectivity and element type rather than on global mesh size, and reported as typically $h = 6.6256 \times 10^{-34}\ \mathrm{J\,s}$9–$\alpha$0 colors in 2D and $\alpha$1–$\alpha$2 in 3D.

The reconstruction workflow is explicitly described as: build the sparsity pattern from mesh connectivity and constraints if present; color the degrees of freedom so that no two within a color share a row in $\alpha$3; for each color, form a seed vector, compute one JVP, and scatter outputs into the corresponding sparse entries; then assemble the sparse matrix. In the presence of Lagrange multipliers, the sparsity pattern is augmented with the pattern of $\alpha$4 obtained by AD.

These mechanisms are presented as the resolution of the standard global-AD trade-off: global differentiation retains expressivity, while batched evaluation, matrix-free operator application, and sparse coloring recover scalability.

6. Problem classes, performance envelope, and implementation workflow

Tatva is described as supporting a broad set of variational and coupled formulations (Pundir et al., 12 Feb 2026). Multi-point constraints can be enforced by condensation, by Lagrange multipliers through a saddle-point Lagrangian

$\alpha$5

or by a penalty contribution

$\alpha$6

Mixed-dimensional coupling includes cohesive interfaces through a surface energy $\alpha$7 with irreversibility enforced by tracking $\alpha$8, and embedded fibers through an added 1D fiber energy coupled to the bulk by precomputed barycentric projections. Neural components are included in two forms: learned constitutive terms $\alpha$9 at quadrature points, stabilized by energy shifting and a convex base term, and a Neural-Operator Element Method in which a subdomain is replaced by a neural operator $10^4$0 added directly to $10^4$1.

The implementation workflow is centered on a user-supplied Python function $10^4$2 over a flattened global degree-of-freedom vector. The framework provides gather-evaluate-reduce primitives through Operator(mesh, element, batchsize), including op.grad(u) for batched gradients at quadrature points, op.eval(u) for batched interpolation, and op.integrate(density) for batched quadrature reduction. Dirichlet and periodic conditions are implemented by wrapping $10^4$3 before evaluating $10^4$4, so that AD of the wrapped energy yields the reduced residual and tangent automatically.

The paper reports strict $10^4$5 scaling in execution time and stable GPU throughput for both matrix-free JVPs and coloring-based sparse differentiation. On an NVIDIA A100 (40 GB), the reported matrix-free JVP throughput is about 47 million DoFs/s in 2D and about 13 million DoFs/s in 3D, essentially constant as DoFs grow. For sparse differentiation, the reported throughput is about 8.7 million DoFs/s in 2D and about 0.5 million DoFs/s in 3D. Linear scaling is reported over three orders of magnitude, up to 10 million DoFs for JVP; explicit sparse differentiation in 3D is reported as memory-limited around 5.1 million DoFs on a single 40 GB GPU. The paper further states that scatter-add assembly shows decreasing throughput with problem size because of memory contention and irregular access, whereas tatva’s throughput remains constant.

The documented examples include a linear plate-with-hole with Dirichlet condensation and agreement with Kirsch’s solution; a hyperelastic plate coupled to a rigid disk via rigid-body condensation; periodic homogenization with Lagrange multipliers and a homogenized tensor $10^4$6 obtained by differentiating the macroscopic stress map with respect to applied strain; 3D Hertz contact using a penalty-based contact energy and two-pass node-to-surface gap detection; mixed-dimensional fracture with cohesive interfaces and dissipated energy matching $10^4$7; embedded fibers tied by barycentric mapping; thermoelasticity–phase-field fracture–transient heat coupling; and a non-variational advection–diffusion problem on a sphere handled through a virtual-work functional $10^4$8 with verified mass conservation (Pundir et al., 12 Feb 2026).

The principal limitations identified in the paper concern explicit sparse differentiation on a single device, where compiler graph size limits 3D sparse assembly near 5.1 million DoFs on 40 GB GPUs, and problems with evolving connectivity or highly nonlocal interactions, where a priori sparsity becomes more difficult. Future directions named in the paper include sparsity detection from probing, advanced coloring such as star or acyclic coloring, multi-device sharding of batches via JAX distributed mechanisms, and integration with external solvers and preconditioners such as PETSc and AMG. Code and documentation are listed at the project repository and documentation sites, with Python/JAX/XLA as the environment and GPU execution recommended for large problems (Pundir et al., 12 Feb 2026).

7. Comparative significance of the two meanings

Across the two papers, Tatva names two substantially different kinds of explanatory ambition. In the historical-philosophical sense, it refers to elemental principles used to organize the world qualitatively; in the finite-element sense, it names a computational framework in which all governing physics is encoded in one differentiable scalar functional. The former belongs to the pre-empirical and early conceptual search for constituents; the latter belongs to contemporary differentiable scientific computing.

The contrast is especially sharp in methodology. The historical Tatva of Panchmahabhootas does not prescribe a testing protocol, whereas modern particle physics, as described in (Godbole, 2010), proceeds through pattern recognition, scattering as a microscope, and quantitative theory. The computational tatva framework, as described in (Pundir et al., 12 Feb 2026), is even more formalized: physics is specified as $10^4$9, derivatives are generated by AD, and scalability is recovered through JVP-based operator application and coloring-based sparse differentiation.

At the same time, there is a limited conceptual continuity. Both usages are concerned with reducing apparent complexity to a smaller set of organizing principles. This suggests a common abstract theme—seeking compact generative descriptions—while preserving the decisive difference that modern scientific and computational usages require explicit mathematics, reproducible workflows, and performance characteristics that can be benchmarked, validated, and extended.