Los distintos procesos involucrados en la industria química deben ser estudiados cuidadosamente con el fin de obtener productos de calidad al menor costo y causando el mínimo daño al medio ambiente (ej. industria de polímeros sintéticos y diseño racional de fármacos). Hace ya varios años que distintos métodos computacionales son utilizados en la industria química con el fin de lograr esos objetivos. En particular, el modelado QSAR/QSPR es una técnica de gran interés dentro del área de la informática molecular, ya que permite correlacionar de manera cuantitativa características estructurales de una entidad química con una determinada propiedad físico-química o actividad biológica. El objetivo de esa tesis fue desarrollar distintas metodologías para asistir a expertos en informática molecular en el proceso de predicción de propiedades fisicoquímicas o de actividad biológica. Más específicamente, las técnicas desarrolladas se enfocan en incorporar al proceso de modelado predictivo QSAR/QSPR, el conocimiento del experto en el dominio. De esta manera se logran mejorar ciertas características de los modelos, tales como su interpretación en términos físicos-químicos, las cuales permite aumentar la generalidad del modelo. Al respecto, se ha implementado una herramienta de analítica visual, denominada VIDEAN, que combina métodos estadísticos con visualizaciones interactivas para elegir un conjunto de descriptores que predigan una determinada propiedad objetivo. Otro de los aportes de esta tesis está relacionado con el dominio de aplicación de un modelo QSAR/QSPR. En este sentido, se ha implementado una técnica para determinar el dominio de aplicación de modelos de clasificación. Esto representa una novedad dado que la mayoría de las técnicas desarrolladas para este fin apuntan exclusivamente a los modelos de regresión. Los métodos implementados han sido evaluados mediante el estudio de propiedades de relevancia para tres campos de aplicación: el diseño racional de fármacos, el diseño de materiales poliméricos (plásticos) y las ciencias ambientales. Con este fin, se han desarrollado numerosos modelos predictivos de regresión y clasificación. En el área de diseño racional de fármacos, las propiedades que se estudiaron están relacionadas con el comportamiento ADMET (absorción, distribución, metabolismo, excreción y toxicidad) de los mismos: absorción intestinal humana (Human Intestinal Absorption, HIA) y el pasaje de la barrera hemato-encefálica (Blood-Brain Barrier, BBB), ambas esenciales para el desarrollo de nuevos fármacos. En el campo de los materiales poliméricos, se exploraron varias propiedades mecánicas, que proporcionan información relacionada con la ductilidad, resistencia y rigidez del material polimérico; y que, junto con otras propiedades, definen su perfil de aplicación estructural. Estas propiedades son: elongación a la rotura (elongation at break), resistencia a tensión en la rotura (tensile strength at break) y módulo elástico (tensile modulus). En el área de medioambiente, la propiedad que se estudió fue el coeficiente de distribución sangre-hígado (log Pliver) en compuestos orgánicos volátiles (VOCs), que son gases que se emiten de ciertos sólidos o líquidos y que son ampliamente utilizados como ingredientes en productos para el hogar (pinturas, los barnices, productos de limpieza, desinfección, cosmética, entre otros). Los resultados de estudios de este tipo de propiedades brindan un panorama de cómo se distribuyen estos tipos de compuestos en el organismo y pueden emplearse para la evaluación de riesgos y toma de decisiones en materia de salud pública. / The various processes involved in the chemical industry must be carefully studied in order to obtain quality products at the lowest cost and causing the least damage to the environment (e.g. synthetic polymer industry and rational drug design). During the last two decades, different computational methods have been used in the chemical industry in order to achieve these objectives. In particular, QSAR/QSPR modeling is a technique of great interest in the area of molecular informatics, since it allows to quantitatively correlate structural characteristics of a chemical entity with a given physical-chemical or biological activity. The objective of this thesis was to develop different methodologies to assist molecular computing experts in the process of predicting physicochemical or biological activity properties. More specifically, the techniques developed focus on incorporating domain expert's knowledge into the traditional automated predictive modeling process. In this way, certain characteristics of the models can be improved, such as their interpretation in physical-chemical terms, which allow to increase the generality on the model. In this sense, a visual analytics tool, called VIDEAN, has been implemented to combine statistical methods with interactive visualizations to choose a set of molecular descriptors that predict a specific target property. Another contribution of this thesis focuses on the implementation of a technique to determine the applicability domain of QSAR/QSPR classification models. In this regard, a technique has been implemented to determine the applicability domain of classification models. This represents a novelty given that most of the techniques developed for this purpose are exclusively intended for regression models. Implemented methods have been evaluated using target properties of relevance in three application areas: rational drug design, design of polymeric materials (plastics) and environmental sciences. To this end, different predictive regression and classification models were proposed that overcome in performance and interpretability to other traditional models have been developed. To this end, numerous regression and classification models have been developed. In rational drug design, the properties that were studied are related to the ADMET behavior (absorption, distribution, metabolism, excretion and toxicity): Human Intestinal Absorption (HIA) and Blood-brain barrier (BBB), both essential for the development of new drugs. In the field of polymeric materials, various mechanical properties, which provide information related to the ductility, strength and rigidity of the polymeric material were explored, and which, along with other properties define its structural application profile. These properties are: elongation at break, tensile strength at break and tensile modulus. In environment area, the property studied was the blood - liver distribution coefficient (log Pliver) in volatile organic compounds (VOCs), which are gases that are emitted from certain solids or liquids and are widely used as ingredients in products for the home (paints, varnishes, cleaning products, disinfection, cosmetics, among others). The results obtained from this studies provide an overview of how these types of compounds are distributed in the body and can be used for risk assessment and public health decision making.

Cette thèse vise à développer une approche basée sur le concept de Graphe Condensé de Réaction (GCR) capable de (i) sélectionner un espace optimal de descripteurs séparant au mieux différentes classes de réactions, et (ii) de préparer de nouveaux descripteurs pour la modélisation « structure–réactivité ». Cette méthodologie a été appliquée à la recherche par similarité dans une base de données contenant 8 classes de réaction différentes; et à la cartographie de son espace chimique en utilisant des cartes de Kohonen et de cartes topographiques génératives. La seconde partie de la thèse porte sur le développement de modèles prédictifs pour le pKa et pour des conditions optimales pour différents types de réaction de Michael impliquant à la fois les descripteurs d’effet électronique et des descripteurs calculés sur les GCR. / This thesis aims to develop an approach based on the Condensed Graph of Reaction (CGR) method able to (i) select an optimal descriptor space the best separating different reaction classes, and (ii) to prepare special descriptors to be used in obtaining predictive structure-reactivity models. This methodology has been applied to similarity search studies in a database containing 8 different reaction classes, and to visualization of its chemical space using Kohonen maps and Generative Topographic Mapping. Another part of the thesis concerns development of predictive models for pKa and for optimal conditions for different types of Michael reaction involving both CGR-based and Electronic Effect Descriptors.

Une approche intégrée physicochimie et relations structure-activité a été mise en œuvre afin d’étudier le phénomène rétention-libération des composés d’arôme dans un gel laitier allégé additionné de pectine. Notre objectif était d’identifier les propriétés moléculaires qui régissent ce phénomène en supposant que la modification de la structure entraîne forcement un changement dans la rétention-libération des composés d’arôme. Dans ce but, nous avons déterminé les coefficients de partage de 28 composés d’arôme dans l’eau, dans des gels de pectine et dans des gels laitiers avec ou sans de pectine, à l’équilibre en utilisant la méthode PRV (Phase Ratio Variation). Nous avons ensuite effectué une étude des relations structure-rétention en évaluant les corrélations entre les coefficients de partage et quatre descripteurs traduisant quatre propriétés moléculaires : l’hydrophobie globale, la surface moléculaire, la polarisabilité et la densité de charge négative. Notre démarche d’étude des relations structure-activité (Structure-Activity Relationships, SAR) consistait à étudier des composés d’arôme appartenant à une gamme de structures variée, dans un même ensemble, puis en sous-groupes en fonction d’une particularité structurale donnée afin de révéler les particularités de la structure qui influent sur le phénomène rétention-libération. La comparaison des rétentions entre les milieux n’a pas montré l’existence d’un effet pectine. Les études des relations structure-activité ont montré l’impact de certaines particularités structurales telles que la ramification et la double liaison sur la rétention. Elles ont également montré que l’hydrophobie globale des molécules n’était pas la propriété moléculaire la plus à même d’expliquer les phénomènes impliqués dans les interactions de molécules odorantes avec les constituants du milieu (eau ou gel laitier). La surface et la polarisabilité rendent mieux compte des rétentions des composés d’arôme. Les corrélations impliquant la surface, la polarisabilité et l’hydrophobie globale, confirment que les interactions de type van der Waals (essentiellement Keesom et London) sont favorables à la rétention dans les gels laitiers et défavorables à la rétention dans l’eau. De même, les corrélations impliquant la densité de charge montrent que les interactions polaires sont favorables à la rétention dans l’eau. Notre choix de départ, qui consistait à faire varier la structure des composés d’arôme afin d’apprécier son effet sur le phénomène rétention-libération des composés d’arôme, s’est avéré concluant, et le groupe de 28 composés permet effectivement de mener une étude quantitative des relations structure-propriété. Cette démarche QSAR pourra se transposer à des systèmes alimentaires simples ou complexes. / An integrated approach physicochemistry and structures activity relationships has been carried out to study the aroma compounds retention-release phenomenon in a fat free dairy gel added with pectin. This study aimed to identify the molecular properties that govern this phenomenon assuming that modifying the structure leads automatically to a change in the retention-release of aroma compounds. For this purpose, we have determined the partition coefficients of 28 aroma compounds in water, in pectin gels and in dairy gels supplemented or not supplemented with pectin, at equilibrium conditions using the PRV method (Phase Ratio Variation). Then, we have performed a structure-retention relationships study for the aroma compounds by estimating correlations between the partition coefficients and four descriptors representing four molecular properties: Global hydrophobicity, molecular area, polarizability and negative charge density. Our methodology concerning the structure-activity relationships study (SAR) consisted on studying a varied range of aroma compounds in terms of molecular structure, first taking into account all of them in the same set, then in separated subgroups according to a given structural particularity in order to reveal which structural particularities control the retention-release phenomenon. The comparison of retention between the several media has not shown any effect of pectin. Structure-activity relationships studies have shown the impact of some structural particularities like branching and double linking. They have also shown that the global hydrophobicity was not the best molecular property to explain the phenomena involved in the interactions between aroma compounds and matrix components (water and dairy gel). Molecular area and polarizability are more likely to report of aroma retention-release. Correlations implying molecular area, polarizability or global hydrophobicity confirm that van der Waals (especially Keesom and London) are involved in the retention in dairy gels and unfavourable in the retention in water. Correlations implying negative charge density show that polar interactions are favourable in the retention in water as well. Our strategy which consisted on varying the structure of aroma compounds to exanimate its effect on the retention-release phenomenon was found to be effective, and the set of 28 aroma compounds allowed as leading a quantitative structure-property relationships study. This QSAR approach can be transposed to simple or complex food systems.

A knowledge of thermophysical properties is necessary for the design of all process units. Reliable property prediction methods are essential because reliable experimental data are often not available due to concerns about measurement difficulty, cost, scarcity, safety, or environment. In particular, there is a lack of prediction methods for solid properties. Predicted property values can also be used to fill holes in property databases to understand more fully compound characteristics. This work is a comprehensive analysis of the prediction methods available for five commonly needed solid properties. Where satisfactory methods are available, recommendations are made; where methods are unsatisfactory in scope or accuracy, improvements have been made or new methods have been developed. In the latter case, the following general scheme has been used to develop correlations: extraction of a training set of experimental data of a specific accuracy from the DIPPR 801 database, selection of a class of equations to use in the correlation, refinement of the form of the equation through least squares regression, selection of the chemical groups and/or molecular descriptors to be used as independent variables, calculation of coefficient values using the training set, addition of groups where refinement is needed, and a final testing of the resultant correlation against an independent test set of experimental data. Two new methods for predicting crystalline heat capacity were created. The first is a simple power law method (PL) that uses first-order functional groups. The second is derived as a modification of the Einstein-Debye canonical partition function (PF) that uses the same groups as the PL method with other descriptors to account for molecule size and multiple halogens. The PL method is intended for the temperature range of 50 to 250 K; the PF method is intended for temperatures above 250 K. Both the PL and PF methods have been assigned an uncertainty of 13% in their preferred temperature ranges based on comparisons to experimental data. A method for estimating heat of sublimation at the triple point was created using the same groups as used in the heat capacity PF method (estimated to have an error of 13%). This method can be used in conjunction with the Clausius-Clapeyron equation to predict solid vapor pressure. Errors in predicted solid vapor pressures averaged about 44.9%. As most solid vapor pressures are extremely small, on the order of one Pascal, this error is small on an absolute scale. An improvement was developed for an existing DIPPR correlation between solid and liquid densities at the triple point. The new correlation improves the prediction of solid density at the triple point and permits calculation of solid densities over a wide range of temperatures with an uncertainty of 6.3%. Based on the analysis of melting points performed in this study, Marrero and Gani's method is recommended as the primary method of predicting melting points for organic compounds (deviation from experimental values of 12.5%). This method can be unwieldy due to the large number of groups it employs, so the method of Yalkowsky et al. (13.9% deviation) is given a secondary recommendation due to its broad applicability with few input requirements.

L'objectif de ces travaux était de développer et d'évaluer des modèles quantitatifs structure-propriété (QSPR) pour la prédiction des propriétés explosives des composés nitroaromatiques, en vue d'une utilisation dans un cadre règlementaire, en particulier celui du nouveau règlement européen REACH. Différentes approches méthodologiques (régressions multi-linéaires, PCA, PLS, arbres de décision) ont été utilisées pour mettre en place des modèles pour la prédiction de la chaleur de décomposition. Les descripteurs des modèles ont été sélectionnés dans un jeu étendu de plus de 300 descripteurs (constitutionnels, topologiques, géométriques et quantiques). Deux premiers modèles avec des domaines d'applicabilité définis et des pouvoirs prédictifs importants ont été obtenus. Des modèles pour trois autres propriétés explosives (la température de décomposition, les sensibilités à la décharge électrique et à l'impact) ont ensuite été développés, avec des performances similaires voire supérieures aux modèles existants. Enfin, l'analyse des mécanismes réactionnels sous-jacents, menée à l'aide de la DFT, a permis de mettre en évidence la présence de chemins de décomposition spécifiques au sein des composés nitroaromatiques et a ainsi complété l'approche QSPR en termes d'interprétation phénoménologique. Cette étude a donc pris en compte l'intégralité des principes mis en place par l'OCDE pour la validation des modèles QSAR/QSPR dans un usage règlementaire (cible expérimentale, structure du modèle, validation, domaine d'applicabilité et interprétation des mécanismes sous-jacents). Deux modèles prédictifs ont même été développés pour la chaleur de décomposition des composés nitroaromatiques.

Una de las aplicaciones más importante de los procesos de filtración por membrana es en el área de tratamiento de agua por ultrafiltración, nanofiltración u ósmosis inversa. Entre los problemas más serios encontrados en estos procesos destaca la aparición de los fenómenos de ensuciamiento y envejecimiento de las membranas que limitan la eficacia de la operación tanto en la separación de los solutos, como en el flujo de permeado, afectando también el ciclo de vida de las membranas.Para reducir el coste de la producción y mejorar la robustez y eficacia de estos procesos es imprescindible disponer de modelos capaces de representar y predecir la eficiencia y el comportamiento de las membranas durante la operación. Una alternativa viable a los modelos teóricos, que presentan varias particularidades que dificultan su postulado, la constituyen los modelos basados en el análisis de los datos experimentales, entre cuales destaca el uso de las redes neuronales. Dos metodologías han sido evaluadas e investigadas, una constando en la caracterización de las interacciones entre las membranas y los compuestos orgánicos presentes en el agua de alimentación, y la segunda basada en el modelado de la dinámica de operación de las plantas de desalinización por ósmosis inversa.Relaciones cuantitativas estructura‐propiedad se han derivado usando redes neuronales de tipo back‐propagation, para establecer correlaciones entre los descriptores moleculares de 50 compuestos orgánicos de preocupación para la salud pública y su comportamiento frente a 5 membranas comerciales de ósmosis inversa, en términos de permeación, absorción y rechazo. Para reducir la dimensión del espacio de entrada, y para evitar el uso de la información redundante en el entrenamiento de los modelos, se han usado tres métodos para seleccionar el menor número de los descriptores moleculares relevantes entre un total de 45 que caracterizan cada molécula. Los modelos obtenidos se han validado utilizando un método basado en el balance de materia, aplicado no solo a los 50 compuestos utilizados para el desarrollo de los modelos, sino que también a un conjunto de 143 compuestos orgánicos nuevos. La calidad de los modelos obtenidos es prometedora para la extensión de la presente metodología para disponer de una herramienta comprensiva para entender, determinar y evaluar el comportamiento de los solutos orgánicos en el proceso de ósmosis inversa. Esto serviría también para el diseño de nuevas y más eficaces membranas que se usan en este tipo de procesos.En la segunda parte, se ha desarrollado una metodología para modelar la dinámica de los procesos de ósmosis inversa, usando redes neuronales de tipo backpropagation y Fuzzy ARTMAP y datos experimentales que proceden de una planta de desalinización de agua salobre Los modelos desarrollados son capaces de evaluar los efectos de los parámetros de proceso, la calidad del agua de alimentación y la aparición de los fenómenos de ensuciamiento sobre la dinámica de operación de las plantas de desalinización por osmosis inversa. Se ha demostrado que estos modelos se pueden usar para predecir el funcionamiento del proceso a corto tiempo, permitiendo de esta manera la identificación de posibles problemas de operación debidas a los fenómenos de ensuciamiento y envejecimiento de las membranas. Los resultados obtenidos son prometedores para el desarrollo de estrategias de optimización, monitorización y control de plantas de desalinización de agua salobre. Asimismo, pueden constituir la base del diseño de sistemas de supervisón capaces de predecir y advertir etapas de operación incorrecta del proceso por fallos en el mismo, y actuar en consecuencia para evitar estos inconvenientes. / One of the more serious problems encountered in reverse osmosis (RO) water treatment processes is the occurrence of membrane fouling, which limits both operation efficiency (separation performances, water permeate flux, salt rejection) and membrane life‐time. The development of general deterministic models for studying and predicting the development of fouling in full‐scale reverse osmosis plants is burden due to the complexity and temporal variability of feed composition, diurnal variations, inability to realistically quantify the real‐time variability of feed fouling propensity, lack of understanding of both membrane‐foulants interactions and of the interplay of various fouling mechanisms. A viable alternative to the theoretical approaches is constituted by models developed based on direct analysis of experimental data for predicting process operation performance. In this regard, the use of artificial neural networks (ANN) seems to be a reliable option. Two approaches were considered; one based on characterizing the organic compounds passage through RO membranes, and a second one based on modeling the dynamics of permeate flow and separation performances for a full‐scale RO desalination plant.Organic solute sorption, permeation and rejection by RO membranes from aqueous solutions were studied via artificial neural network based quantitative structure‐property relationships (QSPR) for a set of 50 organic compounds for polyamide and cellulose acetate membranes. The separation performance for the organic molecules was modeled based on available experimental data achieved by radioactivity measurements to determine the solute quantity in feed, permeate and sorbed by the membrane. Solute rejection was determined from a mass balance on the permeated solution volume. ANN based QSPR models were developed for the measured organic sorbed (M) and permeated (P) fractions with the most appropriate set of molecular descriptors and membrane properties selected using three different feature selection methods. Principal component analysis and self‐organizing maps pre‐screening of all 50 organic compounds defined by 45 considered chemical descriptors were used to identify the models applicability domain and chemical similarities between the organic molecules. The ANN‐based QSPRs were validated by means of a mass balance test applied not only to the 50 organic compounds used to develop the models, but also to a set of 143 new compounds. The quality of the QSPR/NN models developed suggests that there is merit in extending the present compound database and extending the present approach to develop a comprehensive tool for assessing organic solute behavior in RO water treatment processes. This would allow also the design and manufacture of new and more performing membranes used in such processes.The dynamics of permeate flow rate and salt passage for a RO brackish water desalination pilot plant were captured by ANN based models. The effects of operating parameters, feed water quality and fouling occurrence over the time evolution of the process performance were successfully modeled by a back‐propagation neural network. In an alternative approach, the prediction of process performance parameters based on previous values was achieved using a Fuzzy ARTMAP analysis. The neural network models built are able to capture changes in RO process performance and can successfully be used for interpolation, as well as for extrapolation prediction, fact that can allow reasonable short time forecasting of the process time evolution. It was shown that using real‐time measurements for various process and feed water quality variables, it is possible to build neural network models that allow better understanding of the onset of fouling. This is very encouraging for further development of optimization and control strategies. The present methodology can be the basis of development of soft sensors able to anticipate process upsets.

Une approche intégrée physicochimie et relations structure-activité a été mise en œuvre afin d'étudier le phénomène rétention-libération des composés d'arôme dans un gel laitier allégé additionné de pectine. Notre objectif était d'identifier les propriétés moléculaires qui régissent ce phénomène en supposant que la modification de la structure entraîne forcement un changement dans la rétention-libération des composés d'arôme. Dans ce but, nous avons déterminé les coefficients de partage de 28 composés d'arôme dans l'eau, dans des gels de pectine et dans des gels laitiers avec ou sans de pectine, à l'équilibre en utilisant la méthode PRV (Phase Ratio Variation). Nous avons ensuite effectué une étude des relations structure-rétention en évaluant les corrélations entre les coefficients de partage et quatre descripteurs traduisant quatre propriétés moléculaires : l'hydrophobie globale, la surface moléculaire, la polarisabilité et la densité de charge négative. Notre démarche d'étude des relations structure-activité (Structure-Activity Relationships, SAR) consistait à étudier des composés d'arôme appartenant à une gamme de structures variée, dans un même ensemble, puis en sous-groupes en fonction d'une particularité structurale donnée afin de révéler les particularités de la structure qui influent sur le phénomène rétention-libération. La comparaison des rétentions entre les milieux n'a pas montré l'existence d'un effet pectine. Les études des relations structure-activité ont montré l'impact de certaines particularités structurales telles que la ramification et la double liaison sur la rétention. Elles ont également montré que l'hydrophobie globale des molécules n'était pas la propriété moléculaire la plus à même d'expliquer les phénomènes impliqués dans les interactions de molécules odorantes avec les constituants du milieu (eau ou gel laitier). La surface et la polarisabilité rendent mieux compte des rétentions des composés d'arôme. Les corrélations impliquant la surface, la polarisabilité et l'hydrophobie globale, confirment que les interactions de type van der Waals (essentiellement Keesom et London) sont favorables à la rétention dans les gels laitiers et défavorables à la rétention dans l'eau. De même, les corrélations impliquant la densité de charge montrent que les interactions polaires sont favorables à la rétention dans l'eau. Notre choix de départ, qui consistait à faire varier la structure des composés d'arôme afin d'apprécier son effet sur le phénomène rétention-libération des composés d'arôme, s'est avéré concluant, et le groupe de 28 composés permet effectivement de mener une étude quantitative des relations structure-propriété. Cette démarche QSAR pourra se transposer à des systèmes alimentaires simples ou complexes.

Als Modellsystem für Oxide mit Perowskitstruktur ist Strontiumtitanat besonders geeignet, um generalisierbare Erkenntnisse über die Auswirkungen von Defekten zu gewinnen und ausgehend davon Struktur-Eigenschafts-Korrelationen zu diskutieren. Durch den Einsatz verschiedener oberflächensensitiver Methoden lässt sich im Ausgangszustand eine erhöhte Konzentration von Liniendefekten an der Oberfläche nachweisen, die sich durch Temperaturbehandlung verkleinert. Die Defektchemie bei hohen Temperaturen wird zur Simulation der elektrischen Leitfähigkeit in Abhängigkeit vom umgebenden Sauerstoff-Partialdruck genutzt. Die Dotierung des oxidischen Halbleitermaterials ist von Eigendefekten abhängig, wobei Sauerstoff-Leerstellen Donatorniveaus bilden und Strontium-Leerstellen Akzeptorcharakter besitzen. Neben der Diffusionsbewegung dieser Eigendefekte bei hohen Temperaturen kann bei niedrigen Temperaturen ein elektrisches Feld deren Umverteilung bewirken. Damit zeigt sich die Leitfähigkeit abhängig von externen elektrischen Feldern, aber auch weitere Eigenschaften sind auf diesem Wege modifizierbar. Im Rahmen der Arbeit werden strukturelle Änderungen, Valenz-Änderungen und veränderte mechanische Eigenschaften nachgewiesen, die jeweils abhängig vom elektrischen Feld schaltbar sind. Schließlich wird das gezielte Ausnutzen struktureller Defekte für Speicherzellen, die den schaltbaren Widerstand von Metall-SrTiO3-Kontakten zur Grundlage haben, vorgestellt. Die Anwendbarkeit des oxidischen Halbleiters als resistives Speicherelement beruht wiederum auf der Kopplung von Sauerstoff-Leerstellen an das elektrische Feld. / Being a model system for oxides with pervovskite-type of structure, strontium titanate can be used to gain generalizable insights into the consequences of defects and to discuss resulting structure-property relationships. By employing different surface sensitive methods, an increased concentration of line defects is found at the surface that reduces on temperature treatment. The defect chemistry at elevated temperatures is used to simulate the electric conductivity depending on the oxygen partial pressure during annealing. Doping of the oxidic semiconductor depends on intrinsic defects, whereby oxygen vacancies form donor states and strontium vacancies have acceptor character. Beside the diffusion movement of these intrinsic defects at elevated temperatures, at low temperatures an electric field may cause their redistribution. Hence, the conductivity becomes dependent on external electric fields but also other properties can be altered in this way. Within this work, structural changes, valence changes and changing mechanical properties are shown to be switchable by the electric field. Finally, the dedicated usage of structural defects is demonstrated on memory cells that employ the switchable resistance of metal-SrTiO3 junctions. The applicability of the oxidic semiconductor as a resistive memory element is again based on the coupling between oxygen vacancies and the electric field.

With the recent advantages of machine learning in cheminformatics, the drug discovery process has been accelerated; providing a high impact in the field of medicine and public health. Molecular property and activity prediction are key elements in the early stages of drug discovery by helping prioritize the experiments and reduce the experimental work. In this thesis, a novel approach for multi-task regression using a text-based Transformer model is introduced and thoroughly explored for training on a number of properties or activities simultaneously. This multi-task regression with Transformer based model is inspired by the field of Natural Language Processing (NLP) which uses prefix tokens to distinguish between each task. In order to investigate our architecture two data categories are used; 133 biological activities from ExCAPE database and three physical chemistry properties from MoleculeNet benchmark datasets. The Transformer model consists of the embedding layer with positional encoding, a number of encoder layers, and a Feedforward Neural Network (FNN) to turn it into a regression problem. The molecules are represented as a string of characters using the Simplified Molecular-Input Line-Entry System (SMILES) which is a ’chemistry language’ with its own syntax. In addition, the effect of Transfer Learning is explored by experimenting with two pretrained Transformer models, pretrained on 1.5 million and on 100 million molecules. The text-base Transformer models are compared with a feature-based Support Vector Regression (SVR) with the Tanimoto kernel where the input molecules are encoded as Extended Connectivity Fingerprint (ECFP), which are calculated features. The results have shown that Transfer Learning is crucial for improving the performance on both property and activity predictions. On bioactivity tasks, the larger pretrained Transformer on 100 million molecules achieved comparable performance to the feature-based SVR model; however, overall SVR performed better on the majority of the bioactivity tasks. On the other hand, on physicochemistry property tasks, the larger pretrained Transformer outperformed SVR on all three tasks. Concluding, the multi-task regression architecture with the prefix token had comparable performance with the traditional feature-based approach on predicting different molecular properties or activities. Lastly, using the larger pretrained models trained on a wide chemical space can play a key role in improving the performance of Transformer models on these tasks.

no / Efficient management of toxicity information as an enterprise asset is increasingly important for the chemical, pharmaceutical, cosmetics and food industries. Many organisations focus on better information organisation and reuse, in an attempt to reduce the costs of testing and manufacturing in the product development phase. Toxicity information is extracted not only from toxicity data but also from predictive models. Accurate and appropriately shared models can bring a number of benefits if we are able to make effective use of existing expertise. Although usage of existing models may provide high-impact insights into the relationships between chemical attributes and specific toxicological effects, they can also be a source of risk for incorrect decisions. Thus, there is a need to provide a framework for efficient model management. To address this gap, this paper introduces a concept of model governance, that is based upon data governance principles. We extend the data governance processes by adding procedures that allow the evaluation of model use and governance for enterprise purposes. The core aspect of model governance is model representation. We propose six rules that form the basis of a model representation schema, called Minimum Information About a QSAR Model Representation (MIAQMR). As a proof-of-concept of our model governance framework we develop a web application called Model and Data Farm (MADFARM), in which models are described by the MIAQMR-ML markup language. (C) 2013 Elsevier Ltd. All rights reserved.