Simulating Metallic Glass Formation with Orb-v3 is a post about running molecular dynamics simulations to study how a Cu-Zr alloy forms a metallic glass. The author uses a 64% Cu and 36% Zr composition, an (10,10,10) supercell, and the orb-v3-direct-20-omat calculator to push speed and scale. The workflow includes equilibrating a melted alloy at high temperature, then rapid quenching from 2000 K to 300 K at various rates to compare glass formation versus crystallization. The write-up explains key concepts like what glass is in atomic terms, the difference between crystalline order and amorphous structure, and how RDF and coordination numbers help analyze results. It also notes the challenges of achieving crystallization in MD due to time scales and suggests exploring different cooling rates and compositions in future runs. The post includes example data and 3D visualization references to support the findings.
After reading latest paper from Orbital on Orb-v3, I was inspired to try out some larger-scale MD simulations. This first I want to try out is a simulation that shows the formation of a metallic glass from a melted Copper Zirconium alloy. Claude mentioned this compound forms glass easily and is well studied.
Specifically, we're going for 64% CU, 36% Zr. In future simulations, we may try out different ratios as well as different cooling rates.
First, a quick refresher on glasses.
Crystalline solid: Atoms sit on a regular, repeating lattice. A perfect grid - every atom knows exactly where its neighbors are, and this pattern repeats infinitely. FCC copper, BCC iron, diamond structure silicon - all crystals.
Glass (amorphous solid): Atoms are "frozen" in a disordered arrangement - basically a snapshot of the liquid structure.
When you cool a liquid slowly, atoms have time to rearrange into the lowest-energy crystal structure. But if you cool fast enough, atoms don't have time to find their crystalline positions before the whole thing gets too viscous to move. They get stuck in a random arrangement.
Atomic characteristics of glass:
No long-range order (no repeating pattern)
Some short-range order (atoms still prefer certain local coordination)
Continuous random network of bonds
Higher energy than the corresponding crystal (metastable)
Normal glasses (window glass, silica): Made of network formers like SiO₂ - directional covalent bonds make them naturally resist crystallization. Easy to form glasses.
Metals: Usually have non-directional metallic bonding and love to crystallize. They're terrible glass formers. For pure metals, you'd need impossibly fast cooling (~10^15 K/s) to make glass.
Metallic glass (amorphous metal): An alloy engineered to resist crystallization so you can form glass at achievable cooling rates (~10^6 K/s experimentally).
How do you make metal "confused" about crystallizing?
Mix elements with:
Different atomic sizes (hard to pack efficiently)
Multiple competing crystal structures (which one to form?)
Strong chemical interactions (prefers certain local coordination that doesn't match crystal symmetry)
Classic example: Cu-Zr or Pd-Ni-P alloys. The mixture makes it harder for atoms to "agree" on which crystal structure to form before everything freezes up.
Now let's get into the simulation.
Build (10,10,10) supercell, 64:36 ratio, totaling 4000 atoms
Setup calculator, using orb-v3-direct-20-omat for maximum speed and scale
Setup MD environment, Langevin thermostat with friction=0.01 (pretty weak)
Equilibrate melt at 2000 K for 10 ps
10,000 steps, 1 fs step size
Ran in 841 seconds on A100 40GB at ~90% GPU utilization
Rapid quench to form glass, 2000 K to 300 K (room-temperature) in 10 ps
10,000 steps, 1 fs step size
Ran in 891 seconds
Quench rate: 1.70e+14 K/s
This is one run. We repeat the quenching from the same equilibrated melt for varying quench times:
2000 K to 300 K in 20 ps
20,000 steps, 1 fs step size
Ran in 1675 sec
Quench rate: 8.50e+13 K/s
2000 K to 300 K in 60 ps
60,000 steps, 1 fs step size
Quench rate: 2.83e+13 K/s
To provide contrast to these results, we also run the same setup on a 1:1 Cu-Ni alloy melted to 2600 K (originally planned 2200 K).
I think the CuNi material was a little harder to melt. I had to crank it up to 2600 K to properly see full melting. I'm fairly sure if I ran for longer at 2200 or 2000 K it would have melted too, but it was noticeable slower. This is the same mechanism that is the strong crystallization impulse.
This exercise was a good lesson in RDF, coordination number, and understanding those plots.
First, we have the starting supercell. Perfect FCC crystal:
These plots show the RDF and coordination number plots of the first frame of a melt molecular dynamics run. Before the initial velocities are applied to the system, we have a perfect crystal.
After we set the initial velocities at 2600 K from the Maxwell-Boltzman distribution, we lets the system equilibrate for a while.
2600 K is more then enough to melt the system, so it happens quick. This is what the RDF plot looks like after the equilibration. Only a little bit of local order, no long range order.
These plots together describe a system with strong short-range atomic correlations (clear first peak in RDF) and diminishing order beyond a few angstroms, typical of liquids, glasses, or disordered crystalline systems. We see that there is only a significant peak just above 2 Å, matching up with the of Cu and Ni (and more precisely, the typical nearest-neighbor distance).
Just to compare and see why I needed to crank up the temperature, I'll show an example of melt equilibration at a lower temperature for the same amount of time and you'll see there there is still some long-range order.
Here it is at 1800 K melted for 10 ps:
Even at 9 angstroms away, there is still some expectation to find another atom indicating there are still some remnants of the supercell's lattice.
As we can see here (and I could still see visually when I explored the supercell in 3D), there is still some long range order indicating the system has not fully become liquid.
It seems in MD temperature experiments, researchers are far more likely to form glass than proper crystallization of the material. Because of the relatively large timescales needed for proper crystal formation, cooling in MD is likely to be much quicker than would be achievable in a lab.
Of course this is all configurable and there's nothing stopping you from simulating crystallization, provided you have the time to wait and a good MLIP.
I'll say I didn't quite accomplish what I wanted here. I wanted to see some kind of interesting emergent phenomenon in the trajectories, but really didn't see much. Sure there were differences in the phases and different MD setups, but metallic glass doesn't look like much and I have yet to be able to get crystallization to happen in a longer run.
I'll keep trying when I feel like burning more Modal credits but for now I'll end this experiment and move on to some other kinds of simulations.