Researchers at the Simons Foundation have created a simulation of the universe that is both fast and accurate. The new artificial intelligence model used to run the simulation improves upon existing models in such a way that it can offer data on additional parameters even if it is not available in the training data. However, they have faced one problem — explainability.
Explainable AI has always been an obstacle in the way of AI adoption. Even though the common consensus now is that AI as a concept is generally a black box, there is still a large focus on explainable AI. How will it benefit the world at large?
An Introduction To Computational Astrophysics
The model created by the foundation was dubbed the deep density displacement model. It is an AI model that generates complex three-dimensional simulations on a universal scale and does so more accurately that an existing system known as second-order perturbation theory.
To understand the reason behind the creation of the D3M, we must first look into what computational astrophysics stands for. At its base, computational astrophysics focuses on creating an accurate model of the universe to observe any long-term effects of current physical phenomenon. As the scale of space in the area of time extends far beyond human lifespans, universe simulations are used for deriving accurate results.
Most of the modern physics theories come from running complex scientific simulations with variables and replicating the results. Astrophysics’ approach is also not very different — and the accuracy of today’s simulations are integral in ensuring the validity of theories in the future.
The model being used until now was developed in the mid-2000s by researchers at New York University. Known as the 2LPT, the model ran in Fortran77. The 2LPT was an improvement over the existing Zel’dovich approximation method that was used before that and had taken its place as the most accurate universe sim.
The D3M offers a faster result than the 2LPT while being more accurate than the previous model. While this was an expected result, the magnitude of the improvement was not predicted.
The Deep Density Displacement Model
The D3M is a deep neural network model that was built for the express purpose of predicting structure formation in the universe. The main variable that was to be used for this was gravity, as this is most instrumental in determining universe predictions.
The network is backed up by a research paper published in May 2019 titled ‘Learning to predict the cosmological structure formation‘. This paper also proved that the most effective method to simulate the universe was neural networks. According to the abstract:
“It outperforms the traditional fast-analytical approximation and accurately extrapolates far beyond its training data. Our study proves that deep learning is an accurate alternative to the traditional way of generating approximate cosmological simulations.”
Now, astrophysics can be added to the list of fields disrupted by deep learning. Notwithstanding the computational gains exhibited by neural networks, the model could also be used to generate highly complex 3D simulations. This led to deep learning being dubbed as a ‘powerful alternative to traditional numerical simulations’.
Reportedly, the D3N can run simulations in a “few milliseconds”, whereas previous models which were termed ‘fast’ took a couple of minutes at the very least. The slowest and most accurate approach takes hundreds of hours of computation, with D3M completing a simulation with similar accuracy in just under 30 milliseconds. The boost to accuracy was also sizeable, showing the superiority of AI over traditional systems.
The Black Box Problem
The system’s training dataset was simply 8,000 different simulations run on an existing high-accuracy model. Using this training data, the model can not only effectively extrapolate it and create simulations of its own, but also continue to learn from itself and provide more accurate results over time.
The most shocking improvement over existing systems was the ability of the system to accurately consider parameter variations that are not included within its training dataset. This includes variables such as the amount of dark matter in the cosmos and so on. This is a common characteristic of deep learning networks, but to do so on such a scale is unheard of.
An obstacle towards adopting these technologies and extending it to other fields of science is that no one knows how it works. The problem of AI, especially deep learning, being an unexplainable phenomenon, causes problems in observing what makes the model work.
The effectiveness of the model cannot be extended to other similar applications as it is yet to be explained. However, efforts are underway to explain the phenomenon, and may usher in the age of explainable AI.