Quantum simulations happening at scale are beyond human capacity to comprehend

In what seems to be a feud between two opinion groups, the quantum simulation might have opened the doors for a debate on how far AI can generate accurate results, given the scale of calculations it involves requiring millions of predictions that are beyond human comprehension. Recently, Deepmind, Alphabet's AI subsidiary, released a paper," Pushing the frontiers of density functionals by solving the fractional electron problem", concluding with the possibility of simulations at the quantum scale. Eight months later, a batch of Russian and South Korean researchers expressed their apprehension that the conclusion the paper arrived at may be wrong, putting the entire Deepmind's theory in disgrace.

Physical matter is made up of a microcosm of particles numbering in millions, and having zillions of interactions, which implies, that studying the interactions at the quantum level is next to impossible. The complexity increases as one keep adding particles and thus the difficulty in predicting the energy level the electron ends up at.  The researchers are of opinion that the newfound model can help find ways to manipulate the building blocks of matter through artificial intelligence. In the current paper, researchers have claimed to radically improve the functions – earlier which were only like guidelines – by developing neural networks to predict the quantum behavior of electrons and other atomic particles. A statement from Deep Mind's blog reads, "By expressing the function as a neural network and incorporating these exact properties into the training data, we learn functionals free from important systematic errors – resulting in a better description of a broad class of chemical reactions."

But why the conclusions cannot hold? Apparently, Russians are not happy with the training procedure, which clearly lacks transparency, because Russian scientists could not understand how a neural network could come to that conclusion. They suspect the credibility of the function of a neural network in finding the probability of any electron at each position as in DFT (Density Functional Theory) proposed way back in the '60s winning the scientists a Nobel Prize.  The commenting researchers since the Fractional Charge, Fractional Spin system explanation is inadequate for coming to such a revolutionary conclusion. They accuse Deepmind's researchers of teaching the neural network to memorize the answers to specific questions that are generally asked in the benchmarking process. The research boasted improvements over the earlier model DM21m, a neural network model for mapping electron density to chemical interaction energy. However, the Russians couldn't find a viable reason why the improved model DM21 is superior. In the comment paper, they stated, "In our opinion, the improvements in the performance of DM21 on the BBB test dataset relative to DM21m may be caused by a much more prosaic reason: an unintended overlap between the training and test datasets."

In a quick and ready retort, posted on the same day as the comment, DeepMind explained the authenticity of the improvised neural network DM21, pointing to the technical details and saying, "We disagree with their analysis and believe that the points raised are either incorrect or not relevant to the main conclusions of the paper and the assessment of the general quality of DM21." While the Russian response is awaited, we are left at an interesting juncture – Can AI be trusted with its predictions – particularly when corporate bigwigs with commercial interests at stake, are involved in its development?  Will AI has to stay under the umbrella of secrecy so as to garner industry support? Open sourcing though seems to be a discouraging factor for individual interests, it might be a blessing in disguise in long term, giving way to higher usability and strong interoperability of AI/ML models.