Artificial intelligence is not yet capable of effectively solving all the tasks that arise in scientific data processing. One of the most difficult remains the real-time reconstruction of particle tracks. This was stated by Vladimir Korenkov, scientific director of the Laboratory of Information Technologies at JINR, speaking during the round table “AI in Science and Industry: Challenges and Opportunities,” organized as part of JINR Days in Armenia.

This refers to the task in which, based on signals recorded by a detector, it is necessary to reconstruct particle trajectories and determine the parameters of their motion. In modern elementary particle physics experiments, this is one of the key data processing procedures: trackers record numerous signals as charged particles pass through them, after which computational algorithms reconstruct the corresponding trajectories.

Real time remains the main constraint

According to Korenkov, it is precisely the need to recognize tracks in real time that significantly complicates the application of AI. To solve this task, specialists use machine learning and deep learning, but so far they have been unable to achieve the system speed required for real-time operation.

One solution, he said, is to divide the process into several stages. Training neural networks remains a separate problem: the model has to be retrained almost constantly, since the very process researchers work with is, as Korenkov describes it, insufficiently deterministic and more difficult to reproduce consistently.

Thus, the limitation is not the lack of machine learning methods themselves, but the need to adapt them to changing task conditions while simultaneously ensuring the required processing speed.

The problem is of fundamental importance for modern high-energy physics. The volumes of data coming from experimental facilities are constantly growing, and some of the information must be processed directly during the experiment.

Data analysis remains a more predictable task

According to Korenkov, the situation is significantly simpler in the field of data analysis. There, in his assessment, the processes are more predictable, so the artificial intelligence methods being used already work much more effectively.

At the same time, the Laboratory of Information Technologies at JINR is not limited to using any single algorithm. Korenkov noted the use of hybrid approaches, in which dozens of different methods are used simultaneously to solve a single task. Specialists combine them to identify the most effective elements of each approach.

Korenkov believes that the possibilities for applying artificial intelligence will expand as the methods themselves develop. However, there is still no universal solution for all data processing tasks.

According to him, specialists are gradually moving closer to solving the most difficult tasks through the development of algorithms and the combination of various methods. At the same time, the growth of data volumes in itself is not an insurmountable problem for computing infrastructure: in the future, this already means working with exabyte-scale data.

The experience of the Laboratory of Information Technologies at JINR shows a dual situation: artificial intelligence is becoming an increasingly important tool for scientific computing, but its application depends on the nature of the specific task. Where data and processes lend themselves to stable modeling, AI is already demonstrating practical effectiveness. In tasks requiring rapid adaptation and real-time information processing, technological limitations remain.