Deep Learning for f (T) Gravity: Joint PINN Reconstruction and MCMC Analysis of Data Sets

For decades, astronomers have watched the universe expand at an ever-increasing pace. This discovery, made in the late 1990s, upended the long-held belief that gravity should be slowing the cosmos down. To explain this acceleration, scientists proposed the existence of a mysterious force called dark energy, often modeled as a constant energy filling empty space. However, this standard explanation, known as the Lambda Cold Dark Matter model, faces deep theoretical puzzles. It leaves physicists wondering why the universe is accelerating now, and it offers no fundamental reason for the specific strength of this mysterious push. Because of these gaps, researchers have begun exploring whether gravity itself might behave differently on the largest scales, perhaps changing its rules as the universe stretches. One such idea is a theory called teleparallel gravity, which describes the force of gravity not as the bending of space, but as a twisting, or torsion, of the fabric of the cosmos.

A team of researchers from India has taken a fresh approach to testing these ideas. Instead of guessing what the rules of this twisting gravity might look like, they built a computer system designed to learn the rules directly from the data. They combined two massive collections of astronomical observations: one that measures how fast the universe is expanding at different times in its history, and another that tracks the brightness of exploding stars to map out cosmic distances. By feeding this data into a specialized type of artificial intelligence known as a physics-informed neural network, the researchers allowed the computer to reconstruct the history of the universe's expansion while strictly obeying the laws of physics. The system did not assume a specific shape for the gravitational force; it simply let the observations guide the shape of the theory.

The result was a smooth, continuous picture of how the universe has expanded over billions of years. From this reconstructed history, the team derived the specific mathematical form of the twisting gravity that would be required to produce such an expansion. They then simplified this complex, data-driven shape into a manageable formula using a cubic polynomial, a type of equation that can curve and twist to fit complex patterns. To ensure their findings were robust, they used advanced statistical methods to calculate the most likely values for the numbers in this equation and to map out the range of uncertainty. The analysis showed that the reconstructed model fits the observational data well, tracking the expansion history closely across the entire range of redshifts measured.

When the researchers compared their new, data-driven model against the standard theory of gravity, the results were nuanced. The standard model, which assumes a constant cosmological constant, still provided a slightly better statistical fit to the data, requiring fewer assumptions to explain the observations. However, the new model did not fail; it remained a viable description of the universe that deviated from the standard theory in a smooth, systematic way. The study demonstrated that mild changes to the laws of gravity are still allowed by current observations. The most significant achievement of this work was not necessarily proving that the standard model is wrong, but rather demonstrating a powerful new method. By combining deep learning with statistical analysis, the researchers created a flexible pipeline that can reconstruct gravitational theories directly from the sky, offering a way to test modified gravity theories against future, more precise data without being trapped by preconceived ideas.