Chemical computers are becoming more and more real

With the right “learning” strategy, even a relatively simple chemical system can perform non-trivial operations. Chemical computers are becoming more and more real, prove scientists at the Institute of Physical Chemistry of the Polish Academy of Sciences in Warsaw.

Wspóhe modern computers use electronic signals, or physical phenomena related to the flow of charge, for calculationsóin electric. However, information can be processed in many ways. For some time the world has been próTo use a signal for this purposeóin chemical. Emerging chemical computers perform only the simplest logical operations for now.

At the Institute of Physical Chemistry of the Polish Academy of Sciences (IPC PAS) in Warsaw, with the help of computer simulations, it was shown that even uncomplicated and already easy to make sets of droplets, in which theórich oscillating chemical reactions take place, can process information in a useful wayób, e.g. With high accuracy detect the shape of a specific object trójdimensional or correctly classify comócancerous tumors into benign and malignant ones.

– Much of the work currently being done in laboratories is focused wokół construction of chemical counterpartsóIn standard logic gates. We approached the issue differently – mówi Dr. Eng. Konrad Gizynski from the IPC PAS.

– We are studying systems of a dozen-odd drops, in which theórymes propagate chemical signals, and we treat each as a whole, as a kind of neural network. It turns out that such networks, even very simple ones, already after krótkim training can cope well with quite sophisticated issues. For example, our state-of-the-art chip detects without much problem the shape of a sphere in a set of wspóspatial coordinates x, y, z – explains.

How it works?

As the IPC PAS informs in a press release sent to PAP, the systems studied at the Institute work thanks to the Belousov-Zhabotinsky reaction occurring in each of theólnych droplets. The reaction is cyclic: once one cycle is complete, the reactants necessary to start the next cycle are reconstituted in solution. Before the reaction stops, the droplet usually performs tens to hundreds of oscillations. The course of the reaction is easy to observe, since ferroin, which is the catalyst, changes itsój color during the cycle.

In shallow dishes with a thin layer of solution, the effect is spectacular. In the liquid there are color stripes spreading in all directions – chemical fronts. In droplets the fronts can also be seen, but in practice the phase of the cycle is simply indicated by the color of the droplet. When the cycle starts, the droplet rapidly „excites” and turns blue, after which it gradually returns to its initial state, in which theórhyme is red.

– The basis of the operation of our systemsów is the mutual communication between drops: when the drops are in contact, then chemical excitations can transfer from drop to drop. In other words, one droplet can induce a reaction in another droplet. Important here is the fact that the excited droplet cannot be immediately re-excited. MóIn somewhat colloquial terms, before the next excitation it has to momentarily "rest", so that later wrótion to its original state,” explains Dr. Gizynski.

In order for the system to process the information,” explain the specialists of the IPC PAS, “it has to be fed into it, and after processing, it has to be read out. It is also important to be able to controllably modify how the system processes information. For this task, the researchers from the IPC PAS used light and an additional (besides ferroin) catalyst: ruthenium. The ruthenium-catalyzed Belousov-Zhabotinsky reaction has an important feature of the. It is inhibited by blue light, which means that under intense illumination the droplets stop oscillating. Thus, by changing the illumination of a particular droplet, one can decide whether to participate in information processing or not, and excite it according to an arbitrary pattern. In the case of drops used for information input, longer illumination time can then be interpreted as e.g. as a larger value of the input wspóspatial row.

Learning process

– In simulations, in whichóhe Warsaw scientists were looking for methods of detecting the shape of a sphere, square networks of 2×2, 3×3, 4×4 and 5×5 droplets were considered. To represent the actual rate of chemical reactions, it was assumed that the droplet spends one second in the excited state, and to the state in which theórhyme can be stimulated again, it returns after ten seconds,” reads the message sent to PAP.

Before starting the learning process, the researchers still had to create a suitable „manual” with the description of the sphere. For this purpose, a random set of pointów, którym is assigned a value of 1 when the point belonged to the sphere, or a value of 0 when the point lay outside it. The database prepared in this way became the basis of the learning process, in which theóin which each component of the point (x, y, z) determined the illumination time of another input droplet.

To train the droplet system to detect the shape of the sphere, researchers at the IPC PAS used an algorithmóin evolutionary. The learning process began by randomly generating 30 patternsóIn the illumination of the system, in whichórych certain drops were used to input information, while others remained inactive for the time being optimized. After the droplet system processed the entire database, it was checked for whichórej drop evolution is best correlated with the expected outcome of the. The drop was treated as an initial. From such a first-generation system obtainedóThe best few, „multiplied” them by making small changes along the way („mutations”) in ways of illumination and the learning cycle was started at the beginning of the. The learning process continued for 500 generations.

– The best results were achieved for the 4×4 droplet system. It showed the highest efficiency in detecting the shape of the sphere, at 85 percent. In addition, it acquired this ability the fastest, as in just 150 generations of. The 5×5 layout may be mówould be better, but to test it, the learning process would have to be extended over 500 generations – mówi Dr. Gizynski.

A lot of possibilities

Droplet systems do not interpret incoming data, they only look for correlations between them („shapeów”) similar to that in whichórej finding them was trained. Instead of wspóspatial coordinates pointów with the shape of the sphere, so data with a different meaning can be entered into them, such as. related to the róhe different features of the comótumor rec. The móheads inót time to look for „shape” data corresponding to, for example. benign or malignant tumors.

– Indeed, in one of our recent publications, carried out in coóWorking with colleagues at the University of Jena, we showed that a 5×5 droplet array could classify comótumor rks from the CANCER medical database with a precision of up to 97 percent. Using a classical computerów achieve better results, nevertheless there are classifiers on them that work less efficiently. So chemical information processing, while still severely imperfect, is beginning to offer increasingly interesting and useful possibilities, Dr. Gizynski concludes.