Marco Hoffmann

Marco Hoffmann, Hans Hasse, Fabian Jirasek

Although the pure component vapor pressure is one of the most important properties for designing chemical processes, no broadly applicable, sufficiently accurate, and open-source prediction method has been available. To overcome this, we have developed GRAPPA - a hybrid graph neural network for predicting vapor pressures of pure components. GRAPPA enables the prediction of the vapor pressure curve of basically any organic molecule, requiring only the molecular structure as input. The new model consists of three parts: A graph attention network for the message passing step, a pooling function that captures long-range interactions, and a prediction head that yields the component-specific parameters of the Antoine equation, from which the vapor pressure can readily and consistently be calculated for any temperature. We have trained and evaluated GRAPPA on experimental vapor pressure data of almost 25,000 pure components. We found excellent prediction accuracy for unseen components, outperforming state-of-the-art group contribution methods and other machine learning approaches in applicability and accuracy. The trained model and its code are fully disclosed, and GRAPPA is directly applicable via the interactive website ml-prop.mv.rptu.de.

Marco Hoffmann, Thomas Specht, Nicolas Hayer et al.

Machine learning (ML) enables the development of powerful methods for predicting thermophysical properties with unprecedented scope and accuracy. However, technical barriers like cumbersome implementation in established workflows hinder their application in practice. With MLPROP, we provide an interactive web interface for directly applying advanced ML methods to predict thermophysical properties without requiring ML expertise, thereby substantially increasing the accessibility of novel models. MLPROP currently includes models for predicting the vapor pressure of pure components (GRAPPA), activity coefficients and vapor-liquid equilibria in binary mixtures (UNIFAC 2.0, mod. UNIFAC 2.0, and HANNA), and a routine to fit NRTL parameters to the model predictions. MLPROP will be continuously updated and extended and is accessible free of charge via https://ml-prop.mv.rptu.de/. MLPROP removes the barrier to learning and experimenting with new ML-based methods for predicting thermophysical properties. The source code of all models is available as open source, which allows integration into existing workflows.

Marco Hoffmann, Thomas Specht, Quirin Göttl et al.

The excess Gibbs energy plays a central role in chemical engineering and chemistry, providing a basis for modeling the thermodynamic properties of liquid mixtures. Predicting the excess Gibbs energy of multi-component mixtures solely from the molecular structures of their components is a long-standing challenge. In this work, we address this challenge by integrating physical laws as hard constraints within a flexible neural network. The resulting model, HANNA, was trained end-to-end on an extensive experimental dataset for binary mixtures from the Dortmund Data Bank, guaranteeing thermodynamically consistent predictions. A novel surrogate solver developed in this work enabled the inclusion of liquid-liquid equilibrium data in the training process. Furthermore, a geometric projection method was applied to enable robust extrapolations to multi-component mixtures, without requiring additional parameters. We demonstrate that HANNA delivers excellent predictions, clearly outperforming state-of-the-art benchmark methods in accuracy and scope. The trained model and corresponding code are openly available, and an interactive interface is provided on our website, MLPROP.

Rebecca Loubet, Pascal Zittlau, Marco Hoffmann et al.

In this short note, we report and analyze a striking event: OpenAI's large language model o3 has outwitted all students in a university exam on thermodynamics. The thermodynamics exam is a difficult hurdle for most students, where they must show that they have mastered the fundamentals of this important topic. Consequently, the failure rates are very high, A-grades are rare - and they are considered proof of the students' exceptional intellectual abilities. This is because pattern learning does not help in the exam. The problems can only be solved by knowledgeably and creatively combining principles of thermodynamics. We have given our latest thermodynamics exam not only to the students but also to OpenAI's most powerful reasoning model, o3, and have assessed the answers of o3 exactly the same way as those of the students. In zero-shot mode, the model o3 solved all problems correctly, better than all students who took the exam; its overall score was in the range of the best scores we have seen in more than 10,000 similar exams since 1985. This is a turning point: machines now excel in complex tasks, usually taken as proof of human intellectual capabilities. We discuss the consequences this has for the work of engineers and the education of future engineers.

Sebastian Houben, Stephanie Abrecht, Maram Akila et al.

The use of deep neural networks (DNNs) in safety-critical applications like mobile health and autonomous driving is challenging due to numerous model-inherent shortcomings. These shortcomings are diverse and range from a lack of generalization over insufficient interpretability to problems with malicious inputs. Cyber-physical systems employing DNNs are therefore likely to suffer from safety concerns. In recent years, a zoo of state-of-the-art techniques aiming to address these safety concerns has emerged. This work provides a structured and broad overview of them. We first identify categories of insufficiencies to then describe research activities aiming at their detection, quantification, or mitigation. Our paper addresses both machine learning experts and safety engineers: The former ones might profit from the broad range of machine learning topics covered and discussions on limitations of recent methods. The latter ones might gain insights into the specifics of modern ML methods. We moreover hope that our contribution fuels discussions on desiderata for ML systems and strategies on how to propel existing approaches accordingly.

Marco Hoffmann, Daniel Mendez, Fabian Fagerholm et al.

There have been recent calls for research on the human side of software engineering and its impact on various factors such as productivity, developer happiness and project success. An analysis of which challenges in software engineering teams are most frequent is still missing. We aim to provide a starting point for a theory about relevant human challenges and their causes in software engineering. We establish a reusable set of challenges and start out by investigating the effect of team virtualization. Virtual teams often use digital communication and consist of members with different nationalities. We designed a survey instrument and asked respondents to assess the frequency and criticality of a set of challenges, separated in context "within teams" as well as "between teams and clients", compiled from previous empiric work, blog posts and pilot survey feedback. For the team challenges, we asked if mitigation measures were already in place. Respondents were also asked to provide information about their team setup. The survey also measured Schwartz human values. Finally, respondents were asked if there were additional challenges at their workplace. We report on the results obtained from 192 respondents. We present a set of challenges that takes the survey feedback into account and introduce two categories of challenges; "interpersonal" and "intrapersonal". We found no evidence for links between human values and challenges. We found some significant links between the number of distinct nationalities in a team and certain challenges, with less frequent and critical challenges occurring if 2-3 different nationalities were present compared to a team having members of just one nationality or more than three. A higher degree of virtualization seems to increase the frequency of some human challenges.