 So the topic of the presentation is parsley minor releases updates. We made our first release of open force field small molecule force field code name parsley last year. And since then we released the first minor release of parsley in March this year. And the second minor release is a work in progress aiming to be released in May. And so now for this, in this presentation, I'd like to introduce what has been changed in 1.1.0 and what to expect from 1.2.0. So here's a workflow that we use for valence parameter optimization. For valence parameter optimization, we looked at bond stretchings, angle bandings, and also we look at the barrier heights of torsions. And the data that we fit to, we use data from app initial calculations. So for the bond stretchings and angle bandings, which are marked as red colored here in the left top, the bond stretchings and angle bandings are informed by the ab initio optimized geometries and vibrational frequencies. And the torsion parameters are informed by torsion drive, which is basically energy versus torsion profiles of constrained ab initio optimized geometries. And we start from the Smirnoff 99 Frost, Christopher adopted from Ember 99 and Parmet Frost. And we input them with data and use regularized list squares optimization, which is implemented in force balance software and outputs optimized force field. And for the targets we input in the force balance, we first start from molecule sets. For the previous releases, we used roast sets and coverage set and regenerated the optimized geometry data set and torsion drive data sets with certain criteria. And we submit the data to QC Fractal. And once the calculations are done, we download the data set and convert into force balance readable targets. And this is going to be the input of force balance. So moving into 1.1.0 release, our main goal in this release was to fix some chemical perception problems we were aware of in 1.0.0. And we did fix some problems by modifying some parameter definitions and by adding new parameter definitions. And we have three major changes, improvements in 1.1.0. First one is a fix for tetrasol optimization issue. So after our first release of 1.0.0, Ian Craig tested 1.0.0 and found out that it failed to optimize, it failed to reproduce DFT optimized geometries of n-ethyl tetrasol and shared the result with us and we looked into it and were inspecting the problems. And we found out that there was a missing improper torsions and associated torsion, proper torsions which were needed to properly describe the trivalent nitrogen center in the tetrasol. So we added the improper torsions, new improper torsion I3V and new associated torsions T51C and T51CH. And if you see the right-hand side, this is a QM optimized geometry of n-ethyl tetrasol. And you can see from the 1.0.0, we couldn't make a planar tetrasol nitrogen centered area, as you can see in the red-colored structures. But if you see the green-colored structures which were obtained by 1.0, we could successfully reproduce the QM optimized geometries better than the previous release. And secondly, we made some modification for some nitrogen-nitrogen rotations. We're looking into the benchmark results with the test sets. I found out that significant deviations in the M-M optimized geometries from QM optimized geometries, especially for the molecules who's having nitrogen-nitrogen bond in it. So you can see one example here. This is the molecule who's having two aromatic rings connected to the nitrogen chains. And if you see the left-hand side figure, the green-colored structures are the M-M optimized geometry from 1.0.0. And you can see they couldn't reproduce the planar structure around the nitrogen-nitrogen bond. And when I looked into it, the torsion assigned to item 10, 13, 14, 15, or T130, and its periodicity was set to 1. And with the periodicity 1, we couldn't make the planar structure like the QM optimized geometries. So by simply changing the periodicity from 1 to 2, if you see the right-hand side, we could make the planar structure around the nitrogen-nitrogen center properly. And it agreed more with the QM optimized geometries. And lastly, we were aware of that we had some parameters that were overly general. So if you see the right-hand side, the top figure, this is the QM values versus M-M values for the angles whose having its central atom as divalent nitrogen. And you can see above 140, there's small groups which are separated with the rest of the thoughts. So we were thinking that we should separate them into different groups so that we can better describe chemistry. So for this A22, we added new term A22A to properly describe the conjugation effect of a nitrogen-carbon surfer bond. And the result equilibrium angle values were quite different between A22 and A22A, which validate the separations. In that sense, we also added new bond terms B14A and B36A to properly describe the single bond between Sb2 carbon and oxygen with negative one charge and the double bond between nitrogen with positive one charge and the nitrogen with negative one charge. So these were three major improvements that we made in 1.1.0. And moving into 1.2.0 partially where we were focusing mainly on modifying the parameter definition to improve force field in 1.1.0, we were focusing on the design of data sets that we are using in the fitting procedure in 1.2.0. So for this purpose, we generated a second generation optimized geometry data set and second generation portion-drive data sets. And to do this, we added three more molecule sets in the generation procedures with a rush set and copper set that we used before. We added visor discrepancy set, which contains 100 challenging molecules whose geometry, whose DFT optimized geometries were significantly different with OPRS3. And we also added e-molecules discrepancy set, whose geometries, which were significantly different in Semyonov-99 first relative to other force fields, and we also added Bayer sets. And using these five input molecule sets, we generated second generation data sets in a way that it can cover more diverse chemical diversity and also cover to have a better coverage of parameters. And we're now running the optimization with this new data set. And since it's still running, it's running now, the slides that I'm showing from this slide is a preliminary fitting result, which was done with a part of second generation training set. So for the optimization, which is now running, we are using 5,000 optimized geometries and around 1,000 vibrational frequencies and 701 detortions. But for this preliminary fittings, we used around 3,000 optimized geometries and 300 vibrational frequencies and around 600 1D torsions. So we started from the same initial guess that we used in 1.1.0, and using the force balance optimization in 28 steps, we could reach the optimizations. And to see the performance of this 1.2.0 preliminary, we tested them with our test sets. So two types of benchmarks were done. First, we checked the QM versus MM optimized geometries. And secondly, we compared the relative energies between conformers at QM optimized geometries. And as a brief overview of the performance of 1.2.0 preliminary, I checked the final objective function value from the force balance single point calculation. So if the lower the final objective value is the better the force field reproduces QM structures and energetics. So if you see the table here, you can see the objective functional value is lower in 1.2.0 preliminary in both primary and full sets. And to take a closer look at into how they reproduce the QM optimized geometries better than the previous releases, I calculated the weighted RMSE value of WRMSE here, which is weighted root mean square deviation of internal coordinates of MM optimized geometries from QM optimized geometry. And I calculated the change in WRMSE value from initial guess to 1.2.0 preliminary. And if you see the plot here, the negative Y values indicates that the WRMSE value from 1.2.0 preliminary is lower than the one from initial guess, which indicates that it's it better reproduce the QM optimized geometries compared to the initial force field. So we could see like overall improvements in reproducing the QM optimized geometries after the optimization. And also I compare the WRMSE value with 1.2.0 preliminary and 1.1.0, which is the previous release, the latest release that we have. And you can see also if the Y value is lower, is lower than the zero value, that means the 1.2.0 preliminary better predicts the QM optimized geometry than 1.1.0. And one interesting thing that I found from this analysis was that for the significantly improved optimized geometries, which had a change in WRMSE value lower than minus 0.25, they were the molecules who's having phosphonal groups in it. As you can see in the right hand side, this is one example of the molecule with a significant improvement. And you can see the transparent red structure is the optimized geometry from 1.1.0 and the green structure is from 1.0, 2.0 preliminary. And if you focus on the transparent red structures, you can see the hydroxyl hydrogen is located too close to the neighbor oxygen, which forms unwanted strong intramolecular hydrogen bond interaction. So from this 1.2.0 preliminary, we could remove this unwanted intramolecular hydrogen bond, as you can see in this figure. And also to see the reproducibility of relative energies between conformers, I calculated the MM relative energies at the QM optimized geometries and calculated the change in relative MM energies and difference between relative MM energies and relative QM energies and made a distribution plot here. And the way, I used two different ways to calculate the MM relative energies. The left hand side figure is for the left hand side figure, I calculated the MM relative energies by taking the difference of MM energies and the MM energy at the QM minima. And for the second figure, I calculated the MM relative energy by taking the difference of MM energies with the MM energy at the MM minima. And as you can see here in both plot, 1.2.0 preliminary centered more into a zero compared to the 1.1.0 with the lower value of mean absolute deviation value. And also in the second figure, you can see the right hand side tails dies out more quickly than 1.1.0 in 1.2.0 preliminary results. So, those are preliminary results that I have. So, now the optimization for 1.2.0 is running now and we are expecting to see the similar improvement in the 1.2.0 candidate as well like we see in the preliminary result. So, thanks for listening to my talk and please look forward to the next release of version 1.2.0 personally. Thanks.