Optimization of Transferable Site–Site Potentials Using a Combination of Stochastic and Gradient Search Algorithms

Abstract

Discontinuous molecular dynamics (DMD) simulation and thermodynamic perturbation theory (TPT) have been used to study thermodynamic properties for organic compounds. The aim is to infer transferable intermolecular potential models based on correlating the vapor pressure and liquid density. The combination of DMD/TPT generates a straightforward global optimization problem for the attractive potential, instead of facing an iterative optimization−simulation type problem. This global optimization problem is then formulated as a black-box optimization problem and solved using a combination of random recursive search (RRS) and Levenberg−Marquardt (LM) optimization. RRS is suitable for black-box optimization problems since its algorithm is robust to the effect of random noises in the objective function and is advantageous in optimizing the objective function with negligible parameters. LM is efficient local to an optimum with a smooth response surface. The local response surface is shown to be smooth but very flat along valleys with a high degree of coupling between the potential parameters. The algorithm is demonstrated with discretized versions of the Lennard-Jones (LJ) potential and a linear step potential using a database of 231 hydrocarbons, alcohols, aldehydes, amines, amides, esters, ethers, ketones, phenols, sulfides, and thiols. A correspondence is established between the discretized LJ potential and the TraPPE model, demonstrating the manner of improving density estimates and a way of expediting improvement of continuous transferable potentials.

Transferable Intermolecular Potential Models for a Broad Range of Organic Compounds

“…8,9 Dedicated sites were needed for a number of other site types like carbons bonded to a halogen. The parameters for all newly introduced site types are given in Table 1.…”

“…The SPEADMD model applies simplified potential models in combination with thermodynamic perturbation theory to accelerate the molecular simulations and transform the characterization of the attractive contributions to the potentials into an efficient nonlinear regression. 8,9 In a grand sense, we undertake to characterize the intermolecular forces for every molecular interaction in nature. Just as civil engineers develop infrastructure by balancing the forces between the objects of their designs, so must chemical engineers ultimately know all intermolecular forces to design the nanostructure of their devices.…”

Uncertainty quantification and propagation of errors of the Lennard-Jones 12-6 parameters forn-alkanes

“…A four parameter optimization is very difficult primarily due to the strong correlation between the CH 3 and CH 2 parameters. 33 Also, the uncertainty quantification approach outlined in Section IV is considerably more difficult for a four parameter optimization. This is because scanning a 4-dimensional parameter space is more computationally expensive than a pair of 2-dimensional scans.…”

Automated parameterization of intermolecular pair potentials using global optimization techniques

“…Different types of global optimizers were tested by the authors with respect to the given calibration problem. In particular, six different algorithms were selected for comparison: pure random search (PRS), recursive random search (RRS) [38], the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) [39,40], differential evolution (DE) [27], a special taboo search (TS) algorithm [30], and Constrained Optimization using Response Surfaces (CORS) [19,37]. Each of them represents a particular class of global optimizers.…”

“…The method was not chosen to compete with the other methods but to assess their performance. The second is a simple recursive random search (RRS) method, which was successfully used for the optimization of force field parameters but in connection with a less costly simulation technique [38]. It is also based on uniform distributions and box constrained search domains.…”