scholarly journals Hybridizing Physical and Data-driven Prediction Methods for Physicochemical Properties

2021 ◽  
Author(s):  
Fabian Jirasek ◽  
Robert Bamler ◽  
Stephan Mandt

We present a generic way to hybridize physical and data-driven methods for predicting physicochemical properties. The approach ‘distills’ the physical method's predictions into a prior model and combines it with sparse experimental data using Bayesian inference. We apply the new approach to predict activity coefficients at infinite dilution and obtain significant improvements compared to the physical and data-driven baselines and established ensemble methods from the machine learning literature.

2020 ◽  
Vol 56 (82) ◽  
pp. 12407-12410
Author(s):  
Fabian Jirasek ◽  
Robert Bamler ◽  
Stephan Mandt

We present a generic, highly effective approach to combine physical and data-driven prediction methods for physicochemical properties based on Bayesian machine learning and model distillation.


2020 ◽  
Vol 7 (1) ◽  
Author(s):  
Miles L. Timpe ◽  
Maria Han Veiga ◽  
Mischa Knabenhans ◽  
Joachim Stadel ◽  
Stefano Marelli

AbstractIn the late stages of terrestrial planet formation, pairwise collisions between planetary-sized bodies act as the fundamental agent of planet growth. These collisions can lead to either growth or disruption of the bodies involved and are largely responsible for shaping the final characteristics of the planets. Despite their critical role in planet formation, an accurate treatment of collisions has yet to be realized. While semi-analytic methods have been proposed, they remain limited to a narrow set of post-impact properties and have only achieved relatively low accuracies. However, the rise of machine learning and access to increased computing power have enabled novel data-driven approaches. In this work, we show that data-driven emulation techniques are capable of classifying and predicting the outcome of collisions with high accuracy and are generalizable to any quantifiable post-impact quantity. In particular, we focus on the dataset requirements, training pipeline, and classification and regression performance for four distinct data-driven techniques from machine learning (ensemble methods and neural networks) and uncertainty quantification (Gaussian processes and polynomial chaos expansion). We compare these methods to existing analytic and semi-analytic methods. Such data-driven emulators are poised to replace the methods currently used in N-body simulations, while avoiding the cost of direct simulation. This work is based on a new set of 14,856 SPH simulations of pairwise collisions between rotating, differentiated bodies at all possible mutual orientations.


Author(s):  
André M. Quintino ◽  
Davi L. L. N. da Rocha ◽  
Roberto Fonseca Jr. ◽  
Oscar M. H. Rodriguez

Abstract Flow pattern is an important engineering design factor in two-phase flow in the chemical, nuclear and energy industries, given its effects on pressure drop, holdup, and heat and mass transfer. The prediction of two-phase flow patterns through phenomenological models is widely used in both industry and academy. In contrast, as more experimental data become available for gas-liquid flow in pipes, the use of data-driven models to predict flow-pattern transition, such as machine learning, has become more reliable. This type of heuristic modeling has a high demand for experimental data, which may not be available in some industrial applications. As a consequence, it may fail to deliver a sufficiently generalized transition prediction. Incorporation of physics in machine learning is being proposed as an alternative to improve prediction and also to reduce the demand for experimental data. This paper evaluates the use of hybrid-physics-data machine learning to predict gas-liquid flow-pattern transition in pipes. Random forest and artificial neural network are the chosen tools. A database of experiments available in the open literature was collected and is shared in this work. The performance of the proposed hybrid model is compared with phenomenological and data-driven machine learning models through confusion matrices and graphics. The results show improvement in prediction performance even with a low amount of data for training. The study also suggests that graphical comparison of flow-pttern transition boundaries provides better understanding of the performance of the models than the traditional metric


2020 ◽  
Vol 110 (09) ◽  
pp. 624-628
Author(s):  
Maximilian Busch ◽  
Thomas Semm ◽  
Michael Zäh

Industrieroboter werden aufgrund ihres großen Arbeitsraumes zunehmend für die Fräsbearbeitungen großer Werkstücke eingesetzt. Dynamische Instabilitäten während des Prozesses schränken jedoch ihre Produktivität ein. Maschinelle Lernverfahren gewinnen hierbei an Popularität, um Strukturmodelle aus experimentellen Daten abzuleiten. Das Institut für Werkzeugmaschinen und Betriebswissenschaften (iwb) der Technischen Universität München entwickelt in Zuge dessen Methoden, die mit maschinellen Lernverfahren Simulations- und Experimentaldaten verbinden, um dadurch die Strukturdynamik von Fräsrobotern zu modellieren.   Industrial robots are increasingly used for milling applications of large workpieces due to their large working area. However, dynamic instabilities during the process limit their productivity. Thus, machine learning methods are becoming increasingly popular for deriving system models from experimental data. The Institute for Machine Tools and Industrial Management (iwb) at the Technical University of Munich is developing methods to fuse simulation data and experimental data using machine learning methods to model the structural dynamics of milling robots.


Metals ◽  
2019 ◽  
Vol 9 (2) ◽  
pp. 220 ◽  
Author(s):  
Johan Stendal ◽  
Markus Bambach ◽  
Mark Eisentraut ◽  
Irina Sizova ◽  
Sabine Weiß

Data-driven or machine learning approaches are increasingly being used in material science and research. Specifically, machine learning has been implemented in the fields of materials discovery, prediction of phase diagrams and material modelling. In this work, the application of machine learning to the traditional phenomenological flow stress modelling of the titanium aluminide (TiAl) alloy TNM-B1 (Ti-43.5Al-4Nb-1Mo-0.1B) is investigated. Three model types were developed, analyzed and compared; a physics-based phenomenological model (PM) originally developed for steel by Cingara and McQueen, a purely data-driven machine learning model (MLM), and a hybrid model (HM), which uses characteristic points predicted by a learning algorithm as input for the phenomenological model. The same amount of data was used to both fit the PM and train the MLM and HM. The models were analyzed and compared based on the accuracy of their predictions, development and computing time, and their ability to predict on interpolated and extrapolated inputs. The results revealed that for the same amount of experimental data, the MLM was more accurate than the PM. In addition, the MLM was better able to capture the characteristic peak stress in the TNM-B1 the flow curves, and could be developed and computed faster. Furthermore, the MLM was able to make realistic predictions for inputs outside the experimental data used for training. The HM showed comparable accuracy to the PM for the experimental conditions. However, the HM was able to produce a better fit for input conditions outside the training data.


Sign in / Sign up

Export Citation Format

Share Document