Modeling dynamical properties of organic crystals with machine learning potentials

Neuronske mre že se nameću kao vrlo vrijedan alat u znanosti zbog svoje mogućnosti uočavanja uzorka iza vrlo kompliciranih pojava. U ovom radu koristit ćemo neuronske mre že kao potencijale pomoću kojih računamo svojstva molekularnih kristala. Mre že koje koristimo dio su ANAKIN-ME metodologije. Uz ANI potencijale koristili smo i popravku za Van der Waalsove interakcije pomoću D4 modela. Najprije smo relaksacijom jedinične ćelije pomoću ANI1x, ANI2x, ANI1x + D4 te ANI2x + D4 potencijala izračunali teorijski optimalan volumen te usporedili s eksperimentalnim podacima. Nakon zadovoljavajućeg slaganja odlučili smo primijeniti ove potencijale na spoj N’-2-propylidene-4-hydroxybenzohydrazide koji pokazuje vrlo rijetki termoodskočni efekt. U kristalnoj strukturi ima 104 atoma u jediničnoj ćeliji što čini tipične ab initio metode iznimno zahtjevnima. Izračunata je fononska gustoća stanja te Helmholtzova slobodna energija za tri različita polimorfa. Potencijali su uspjeli prepoznati neke karakteristike kristala poput karakterističnih energija vibracija te nekih prijelaza polimorfnih faza, no ipak se još ne mogu u potpunosti zamijeniti ab initio metodama za precizne izračune kod ovako kompleksnih struktura. Ipak, zbog svoje brzine predstavljaju ogroman potencijal za budućnost fizike čvrstog stanja te su predmet daljnjeg istraživanja.Because of the ability to recognize patterns behind complex phenomena, neural networks are starting to make their way into the natural and social sciences. In this work we will use neural networks as potentials with which we can calculate various properties of molecular crystals. More specifically, we will use potentials based on ANAKIN-ME methodology. On top of that, we added Van der Waals interaction with D4 model. Firstly, we optimized unit cells of molecular crystals with the help of ANI1x, ANI2x, ANI1x + D4, and ANI2x + D4 potentials, and calculated volumes which were then compared to the experimental data. After achieving reasonable agreement, we decided to use these potentials in order to calculate vibrational properties and Helmholtz free energy. Compound on which we have done the calculations is N’-2-propylidene-4-hydroxybenzohydrazide. This material is very special since it shows particularly rare thermosalient effect. In its crystal structure, it has 104 atoms making typical ab initio calculations very expensive. We proceeded to calculate phonon density of states and Helmholtz free energy. Although potentials do recognize some important features of this crystal like characteristic energy of vibrations of hydrogen atoms or phase transitions, we conclude that these potentials still cannot completely replace ab initio methods. Nevertheless, application of neural networks in solid state phyisics is a hot topic since they possibly represent extremely valuable tool for the field