Skip to content

Latest commit

 

History

History
59 lines (45 loc) · 2.8 KB

algorithm.md

File metadata and controls

59 lines (45 loc) · 2.8 KB
Raw
  1. Included:

  • anthracene.top, the same top file for all simulations.

  • dg.gro, which can be modified by each run by replacing init-lambda-state with the lambda state to be run at.

  • lambda0X0.gro to initialize simulations at states 0, 10, 20, . . . , 100.

  • A sample analysis script dg.py which takes in all of the .dhdl files, and runs MBAR.

  1. Algorithm would be after each simulation comes in, run the analysis (basic_analysis.py) on the dhdl.xvg simulation outputs that have come in, and decide which lambda the next simulation (or simulations should be run at) should be run at.

  2. Initial algorithm: Identify where the largest estimated uncertainty is between two neighboring states, and run simulations at state that is in between the ones with the largest estimated uncertainty is, i.e. if the largest error is between states 60 and 70, run a simulation at state 65. If there is no state between them, run at one of the two states (whichever has highest uncertainty in the next state over. We will probably have to iterate.

This information is stored in ti.d_delta_f_ and mbar.d_delta_f_, in the off-diagonal elements. Note that ti.d_delta_f_ only shows data at states that HAVE samples, whereas mbar.d_delta_f_ shows data at all states, regardless of whether they have samples. So if you use mbar.d_delta_f_, you don't have to bisect, you could just run a simulation at the states that have the highest uncertainty, which is easier.

Note that if you are running multiple simulations, the states with the lowest uncertainty might be next to each other, so you might want to include some criteria that the next simulation is at least delta lambda = 0.1 from currently running simulations or something, as the two neigboring simulations would be duplicative.

To edit the .mdp for the next run, you just need to fill in the value for init-lambda-state.

To pick which .gro to use, if there does not exist a simulation with the same lambda value as the last .gro. pick whichever one is closest in the LOWER direction (i.e. for 0.69, pick 0.6). This is because the particle is disappearing in the increasing lambda direction, and it's easier to start with a simulation that is at a higher level of insertion.

The length of the current simulation seems reasonable. The analysis completed without error for 1 dhdl.xvg this length. On 4 cores, it took about 10 min for me. per itartion

A sample sample_output.txt of the analysis is included which ran from the .dhdl.xvg files included.

Might be nice to keep the graphs/data over time to see how they evolve over time. We will come up with better ways to show this graphically.

We will come up with better versions to show this eventually.

choosing_lambda.ipynb is a notebook that describes the algorithm to select new states based on the variance measured so far.