Recommended learning plan for melt-quench-recrystallization
Posted: Fri Apr 26, 2024 9:33 am
I have earlier monitored the recrystallization of a binary TMDC from a melt-quenched system using standard Vasp MD using first-principles.
I started with the crystalline system and after randomizing the system at 3000K and then gradually lowered the temperature to just above the melting point of 2000K, followed by a melt-quench to room temperature at a rate of 15K/ps. I then found I could recrystallize the melt-quenched structure in less than a nanosecond of MD by maintaining the temperature just below the melting point. This was an interesting result, but since the calculation used only around 150, I wanted to scale up the size of the calculation using ML. To to this, I attempted to train a MD model within Vasp so I could scale up the calculation. I trained until about 2000 configurations for temperatures between room temperature and 2500K. I then took the initial melt-quenched structure that I mentioned in my first statement and tried to recrystallize it as I had done with the completely first-principles calculation. The result included unphysical bond lengths and was clearly incorrect. To make a long story short, I am looking for advice on what sort of learning procedure that would be effective for learning the force field as well as the number of configurations that are necessary to reproduce the results. Thanks for any advice you can offer.
I started with the crystalline system and after randomizing the system at 3000K and then gradually lowered the temperature to just above the melting point of 2000K, followed by a melt-quench to room temperature at a rate of 15K/ps. I then found I could recrystallize the melt-quenched structure in less than a nanosecond of MD by maintaining the temperature just below the melting point. This was an interesting result, but since the calculation used only around 150, I wanted to scale up the size of the calculation using ML. To to this, I attempted to train a MD model within Vasp so I could scale up the calculation. I trained until about 2000 configurations for temperatures between room temperature and 2500K. I then took the initial melt-quenched structure that I mentioned in my first statement and tried to recrystallize it as I had done with the completely first-principles calculation. The result included unphysical bond lengths and was clearly incorrect. To make a long story short, I am looking for advice on what sort of learning procedure that would be effective for learning the force field as well as the number of configurations that are necessary to reproduce the results. Thanks for any advice you can offer.