#### Want to do dynamics? This is how much it’ll cost.

I’m often asked what’s the biggest system we can simulate with Newton-X. You may guess, there is no definitive answer, as it depends on many variables, like the computational level and the simulation time.

We can, however, make a quick estimate on whether it’s feasible or not to do surface hopping dynamics for a particular system.

Technically, Newton-X doesn’t have any limitation concerning system size. The dynamic allocation of memory takes care of setting up all limits adequately. The main problem is computational cost.

The cost to run excited-state nonadiabatic dynamics with surface hopping depends on:

- how long it takes to compute energy plus energy gradients for a single geometry (let’s call it
*T*in CPU-hours);_{g} - how long the chemical process takes (
*T*in fs);_{p} - how much computational resources you have available.

Based on these quantities, you can do a simple estimate in this way:

To compute one single trajectory with time step Δ*t* (also in fs), you need to compute *N _{p}* =

*T*/Δ

_{p}*t*single geometries.

Therefore, the computer time needed to simulate the trajectory is *T _{t}* =

*N*×

_{p}*T*.

_{g}But you still need to compute *N _{t}* independent trajectories to have statistically relevant results. Then, the total computer time will be

*T*=

_{total}*N*×

_{t}*T*.

_{t}Putting everything together,

*T _{total}* (CPU-hours) =

*N*×

_{t}*T*(CPU-hours) ×

_{p}*T*(fs) / Δ

_{g}*t*(fs).

For instance, suppose i) that it takes *T _{g}* = 1 CPU-hour to compute energy and gradients for a single geometry of your system, ii) that the chemical process you’re investigating takes place within 1 ps (

*T*= 1000 fs), and iii) that you need

_{p}*N*= 100 trajectories. Then, you may have to allocate:

_{t}*T _{total}* = 100 × 1 × 1000 / 0.5 = 200.000 CPU-hours.

I’ve few notes about this estimate:

First, I’m assuming that we have a quantum-chemical electronic-structure method able to get reasonable energies and energy gradients for any geometry visited during dynamics. This isn’t always true. If we use TDDFT, for instance, we won’t be able to describe conical intersections to the ground electronic state. You can know more about this in this paper.

Second, in the example, I took *N _{t}* = 100 trajectories. This is the typical number you find in the papers in this field. Note, however, that the uncertainty associated to

*N*is about ε = 1.96 × √(0.25/

_{t}*N*) for a 95% confidence interval. For

_{t}*N*= 100, this means ε = 0.098 ≅ 10%. If you are interested in a process that occurs with a frequency lower than 10%, you may need many more trajectories.

_{t}Third, in the estimate, I also left out the time for computing the nonadiabatic couplings. This is a fair approximation for couplings based on wave-function overlaps and small double-ζ basis sets. However, a better estimate would be to take *T _{p}* =

*time*(energy) +

*time*(energy gradient) +

*time*(couplings). Depending on the case, the computation of the couplings may take as much time as the computation of the energy gradients.

Well, our dynamics simulations will take 200.000 CPU-hours. This means 23 years in a single-core sequential run! Of course we need to consider parallelization. First, surface hopping is an independent-trajectory approach. It means that each trajectory may be run in different computers. We just put them together for analysis.

Thus, each trajectory in the example takes 2.000 CPU-hours, which means 3 months in single core or roughly 10 days in 8 cores. Therefore, if you have 800 cores available, you may have your simulations done within 10 days.

Finally, the current price of one CPU-hour in one of the French national high-performance clusters is about 0.044 €. This means that 200,000 CPU-hours will cost 8.800 €. Therefore, you must ask yourself: is it really worth doing dynamics?

MB