Fakultät für Bio- und Chemieingenieurwesen

Permanent URI for this collection

http://hdl.handle.net/2003/39256

Browse

Now showing 1 - 9 of 9

Data-based predictive control with nonlinear and multi-step models
(2024) Fiedler, Felix Malte Hermes; Lucia Gil, Sergio; Gros, Sebastian
Over the last decades, model predictive control (MPC) has demonstrated exceptional performance in control tasks across various domains. Unfortunately, this potential is often limited by the challenge of obtaining a control-oriented model. To address this challenge, this thesis explores advanced methods for data-based system identification, specifically focusing on nonlinear state-space identification and linear multi-step identification with deterministic and probabilistic models. The resulting models are used to formulate data-based nonlinear, economic and stochastic MPC controllers. Nonlinear state-space identification with neural networks is proposed for systems with state-feedback and output-feedback. Sampling data for the identification task and formulating the nonlinear MPC controller with neural network system model is enabled by the introduced open-source software do-mpc. Multi-step models predict finite sequences of a dynamic system. For linear systems, multi-step identification boils down to a tractable linear regression task. This thesis proposes an MPC controller with identified multi-step model and compares the approach with the recently popularized data-enabled predictive control method. Additionally, for non-deterministic systems, this thesis introduces a probabilistic multi-step identification approach. Using only recorded data of a noise-affected linear system, the proposed method yields an output-feedback stochastic MPC controller with favorable properties. Furthermore, the inherent advantages of multistep identification over state-space identification for linear systems with measurement noise are demonstrated. It is shown that an identified and recursively evaluated state-space model yields biased multi-step predictions, whereas the respective identified multi-step model is, in expectation, unbiased. A possible extension of the proposed methods is nonlinear probabilistic system identification using neural networks with Bayesian last layer. Another major contribution is the proposal of a novel training algorithm that enhances the predictive distribution of the resulting models. This contribution enables the design of data-based stochastic MPC with neural network models for future work.
Improved uncertainty quantification for neural networks with Bayesian last layer
(2023-11-02) Fiedler, Felix; Lucia, Sergio
Uncertainty quantification is an important task in machine learning - a task in which standard neural networks (NNs) have traditionally not excelled. This can be a limitation for safety-critical applications, where uncertainty-aware methods like Gaussian processes or Bayesian linear regression are often preferred. Bayesian neural networks are an approach to address this limitation. They assume probability distributions for all parameters and yield distributed predictions. However, training and inference are typically intractable and approximations must be employed. A promising approximation is NNs with Bayesian last layer (BLL). They assume distributed weights only in the linear output layer and yield a normally distributed prediction. To approximate the intractable Bayesian neural network, point estimates of the distributed weights in all but the last layer should be obtained by maximizing the marginal likelihood. This has previously been challenging, as the marginal likelihood is expensive to evaluate in this setting. We present a reformulation of the log-marginal likelihood of a NN with BLL which allows for efficient training using backpropagation. Furthermore, we address the challenge of uncertainty quantification for extrapolation points. We provide a metric to quantify the degree of extrapolation and derive a method to improve the uncertainty quantification for these points. Our methods are derived for the multivariate case and demonstrated in a simulation study. In comparison to Bayesian linear regression with fixed features, and a Bayesian neural network trained with variational inference, our proposed method achieves the highest log-predictive density on test data.
Probabilistic multi-step identification with implicit state estimation for stochastic MPC
(2023-10-20) Fiedler, Felix; Lucia, Sergio
Stochastic Model Predictive Control (SMPC) is a promising solution for controlling multivariable systems in the presence of uncertainty. However, a core challenge lies in obtaining a probabilistic system model. Recently, multi-step system identification has been proposed as a solution. Multi-step models simultaneously predict a finite sequence of future states, which traditionally involves recursive evaluation of a state-space model. Particularly in the stochastic context, the recursive evaluation of identified state-space models has several drawbacks, making multi-step models an appealing choice. As a main novelty of this work, we propose a probabilistic multi-step identification method for a linear system with noisy state measurements and unknown process and measurement noise covariances. We show that, in expectation, evaluating the identified multi-step model is equivalent to estimating the initial state distribution and subsequently propagating this distribution using the known system dynamics. Therefore, using only recorded data of an unknown linear system, our proposed method yields a probabilistic multi-step model, including the state estimation task, that can be directly used for SMPC. As an additional novelty, our proposed SMPC formulation considers parametric uncertainties of the identified multi-step model. We demonstrate our method in two simulation studies, showcasing its effectiveness even for a nonlinear system with output feedback.
Synthesis of aspartic proteases probes and their application for interaction identification and binding hotspots mapping
(2023) Chen, Suyuan; Sickmann, Albert; Kaiser, Markus
Aspartic proteases play a crucial role in human physiology and pathologyincluding as biomarkers for breast cancer and Alzheimer's disease, and as potential drug targets for infectious diseases. However, chemical probes for photoaffinity labeling (PAL) of these proteases are underdeveloped. We develop a full on-resin synthesis of clickable PAL probes based on the natural product inhibitor pepstatin, incorporating a minimal diazirine photo-reactive group. The positioning of this group in the inhibitor determines the labeling efficiency. Effective probes sensitively detect cathepsin D, a biomarker for breast cancer, in cell lysates. Through chemical proteomics experiments and deep learning algorithms, we also identify sequestosome-1 as a direct interaction partner and substrate of cathepsin D. PAL combined with tandem mass spectrometry (MSn) can reveal interactions between small molecule drugs and protein in biological environments. However, the direct detection of the ‘hotspots’ of the photo-crosslinking sites by MSn is challenging because of the unexpected fragmentation of small molecule drugs, especially when these are small peptides. We synthesize and introduce sulfoxide diazirine (SODA) building blocks to peptide-like probes for PAL. Those MS-cleavable probes enable a MS2 cleavage event that generates a probe-derived reporter ion and a minimal fragment on the modified peptide. Following a subsequent MS3 fragmentation event, we show that this strategy allows for unbiased identification of modification sites and mapping of binding hotspots of peptide-like bio-active molecules. Overall, our study presents the synthesis of aspartic proteases probes and their application for interaction identification and binding hotspots mapping. However, further improvement is required for this study to achieve broader application. We propose a number of possible follow-up experiments and discuss future prospects in chapter 6.
Cell‐free protein synthesis for the screening of novel azoreductases and their preferred electron donor
(2022-05-20) Rolf, Jascha; Ngo, Anna Christina Reyes; Lütz, Stephan; Tischler, Dirk; Rosenthal, Katrin
Azoreductases are potent biocatalysts for the cleavage of azo bonds. Various gene sequences coding for potential azoreductases are available in databases, but many of their gene products are still uncharacterized. To avoid the laborious heterologous expression in a host organism, we developed a screening approach involving cell-free protein synthesis (CFPS) combined with a colorimetric activity assay, which allows the parallel screening of putative azoreductases in a short time. First, we evaluated different CFPS systems and optimized the synthesis conditions of a model azoreductase. With the findings obtained, 10 azoreductases, half of them undescribed so far, were screened for their ability to degrade the azo dye methyl red. All novel enzymes catalyzed the degradation of methyl red and can therefore be referred to as azoreductases. In addition, all enzymes degraded the more complex and bulkier azo dye Brilliant Black and four of them also showed the ability to reduce p-benzoquinone. NADH was the preferred electron donor for the most enzymes, although the synthetic nicotinamide co-substrate analogue 1-benzyl-1,4-dihydronicotinamide (BNAH) was also accepted by all active azoreductases. This screening approach allows accelerated identification of potential biocatalysts for various applications.
Agar plate-based screening approach for the identification of enzyme-catalyzed oxidations
(2022-09-29) Kinner, Alina; Lütz, Stephan; Rosenthal, Katrin
Biocatalytic oxidation reactions are in high demand. Among the applied enzymes, the heme-thiolate enzyme subfamily of unspecific peroxygenases (UPOs), which relies on hydrogen peroxide as cosubstrate and oxidant, has generated great interest. Almost two decades after their first description, databases provide thousands of putative UPO sequences, but only a few enzymes have been characterized. To address this gap, efficient screening methods for the identification of novel UPOs are necessary. Here, a new screening strategy is presented based on solid cultivation of wild-type fungi. The identification of a promising candidate strain highlights the applicability of this approach.
Comparative life cycle assessment of chemical and biocatalytic 2’3’-cyclic GMP-AMP synthesis
(2022-11-23) Becker, Martin; Ziemińska-Stolarska, Aleksandra; Markowska, Dorota; Lütz, Stephan; Rosenthal, Katrin
Life cycle assessments (LCAs) can provide insights into the environmental impact of production processes. In this study, a comparative LCA was performed for the synthesis of 2’3’-cyclic GMP-AMP (2’3’-cGAMP) in an early development stage. The cyclic dinucleotide (CDN) is of interest for pharmaceutical applications such as cancer immunotherapy. CDNs can be synthesized either by enzymes or chemical catalysis. It is not known which of the routes is more sustainable as both routes have their advantages and disadvantages, such as a poor yield for the chemical synthesis and low titers for the biocatalytic synthesis. The synthesis routes were compared for the production of 200 g 2’3’-cGAMP based on laboratory data to assess the environmental impacts. The biocatalytic synthesis turned out to be superior to the chemical synthesis in all considered categories by at least one magnitude, for example, a global warming potential of 3055.6 kg CO2 equiv. for the enzymatic route and 56454.0 kg CO2 equiv. for the chemical synthesis, which is 18 times higher. This study demonstrates the value of assessment at an early development stage, when the choice between different routes is still possible.
Efficient approximations of model predictive control laws via deep learning
(2023) Karg, Benjamin; Lucia, Sergio; Mesbah, Ali
Model predictive control (MPC) has established itself as the standard method for the control of complex nonlinear systems due to its ability to directly consider constraints and uncertainties while optimizing a control objective. However, the application of MPC requires repeatedly solving an optimal control problem online which can be computationally prohibitive, especially for large systems, for systems with very high control sampling rates and for the implementation on embedded hardware. This thesis presents deep neural networks (DNNs) as a means of enabling the implementation of MPC algorithms when computation power is limited. The expressive capabilities of DNNs are leveraged to closely approximate the control law implicitly defined by the MPC problem. For the online application only evaluating the DNN, an explicit function consisting of simple arithmetic operations, is required. This results in speed-ups of several orders of magnitude in comparison to solving the MPC problem online. Throughout the thesis, we shed light onto various aspects that enable and motivate the usage of DNNs as safe approximate MPC laws. Approaches to modify a DNN after an initial learning phase such that the closed-loop performance is improved are proposed. Further, methods that enable the analysis of the closed-loop behavior to obtain both deterministic and probabilistic guarantees on the online operation regarding safety, performance and stability are presented. The efficacy of the proposed approaches is investigated for a wide range of case studies including a polymerization reactor of industrial complexity. The analysis shows that the DNN controllers do not only outperform other approximate MPC approaches in terms of control performance, memory footprint and evaluation times, but that DNN controllers can even outperform the exact optimization-based MPCs when ideas from reinforcement learning are used. Further it is shown that the DNN controllers can be deployed on embedded hardware such as microcontrollers with small effort.
Probabilistic performance validation of deep learning-based robust NMPC controllers
(2021-07-22) Karg, Benjamin; Alamo, Teodoro; Lucia, Sergio
Solving nonlinear model predictive control problems in real time is still an important challenge despite of recent advances in computing hardware, optimization algorithms and tailored implementations. This challenge is even greater when uncertainty is present due to disturbances, unknown parameters or measurement and estimation errors. To enable the application of advanced control schemes to fast systems and on low-cost embedded hardware, we propose to approximate a robust nonlinear model controller using deep learning and to verify its quality using probabilistic validation techniques. We propose a probabilistic validation technique based on finite families, combined with the idea of generalized maximum and constraint backoff to enable statistically valid conclusions related to general performance indicators. The potential of the proposed approach is demonstrated with simulation results of an uncertain nonlinear system.

Browse

Recent Submissions