With the purpose of promoting the discussion about the running projects, monthly seminars have been held. These meetings comprise an half-hour individual presentation followed by a period of discussion.

Click in the links below to access the summary of the past events:

1st Meeting - "NEURAL NETWORKS - short background and benchmark modelling application"

2nd Meeting - "An Optimized Strategy for Equation-Oriented Global Optimiza

3rd Meeting - "Searching for optimal modelling strategies. Basic concepts and novel approaches

4th Meeting - "On-line Tuning of Neural PID Controller"

5th Meeting - "Crystal Growth of Sucrose � Theoretical and Experimental approach"

6th Meeting - "Industrial Applications of Global Optimization - recent achievements and future perspectives"

7th Meeting - "What is Image Analysis? or How to Make a Computer See"

8th Meeting - "FEUP's Collaboration to the ATLAS project at CERN"

9th Meeting - "Modelling Cells Reaction Kinetics with Modular Artificial Neural Networks"

10th Meeting - "Model Based Design and Optimization of Batch Bioprocesses: Practical Aspects"

11th Meeting - "The Different Phases in the Precipitation of Dicalcium Phosphate Dihydrate.

12th Meeting - "Hybrid Modelling, Optimisation and Control of Fermentation Processes: a General Framework"

13th Meeting - "Data Driven Modeling of Sugar Crystallization for Control"

14th Meeting - "Effect of Solids on Flow Regime Transition in Three-Phase Bubble Columns"

Author: Petia Georgieva
23^rd April 2002

"The talk is going to address recent investigation on application of hybrid (artificial neural network-first principle) modelling to industrial scale (fed-) batch evaporative crystallisation process in cane sugar refining. The hybrid model combines a partial mechanistic model that reflect the general mass, energy and population balances with a neural network to approximate the growth rate, nucleation and agglomeration kinetic parameters. The results obtained demonstrate good agreement between experimental data and hybrid model predictions, which are favourable when compared with a complete mechanistic model. As an introduction, some historical notes on the artificial neural networks and respective influencing publications are also envisaged."

Author: Ricardo Lima
13^th May 2002

"The presentation will be divided in two parts:

I.
As introduction general concepts and definitions are presented, with a special focus in local/global optimums and local/global optimization methods. An integrated strategy of simulation and optimization of steady-state processes, under an equation-oriented environment is presented. This environment uses an optimization strategy to optimize the simulation step, based on the optimum choice of decision variables, and a simulated annealing optimizer to perform the global optimization. By optimizing the simulation step, more accurate results were obtained, as well as significantly reduced CPU times, in comparison with simultaneous solution strategies.

II.
Display of a poster to be presented in ESCAPE 12 at the end of May."

Author: Peter Ho
11^th June 2002

"A range of different procedures for the selection, optimisation and evaluation of models will be discussed, aiming at generating a discussion on the reliability and appropriateness of using certain methodologies for model building. It is envisaged that this discussion will aid in developing "optimal" modelling strategies for building better models for engineering and science problems. Modelling strategies will be discussed under the following headings: Model selection, error assessment, validation, performance, generalization, feature/variable selection and parameter optimisation."

Author: Anton Andrasik
25^th November 2002

"The main topic of the presentation is the development of a methodThe main topic of the presentation is the development of a method for nonlinear control of chemical processes based on on-line training of a neural controller. The proposed method utilizes the well-known gradient descent, where the learning rate is adapted in each iteration step in order to accelerate the speed of convergence. It is shown that the selection of the learning rate results in stable training in the sense of Lyapunov. To demonstrate the feasibility and the performance of this control scheme a continuous-flow stirred biochemical reactor model have been chosen as simulation case study. Simulation results demonstrate the usefulness and the robustness of the control system proposed."

Authors: Emilia Zettergren and Pedro Martins
8^th January 2003

"The influence of hydrodynamic conditions in crystallization will be discussed from a theoretical and experimental point of view. Relations between empirical and theoretical crystal growth models, as well as, quantitative ways to measure the degree of diffusion control will be suggested. Single crystal growth experiments have been carried out with the purpose of testing existing and proposed theories. Practical aspects of the laboratorial work will be summarized."

Author: Romualdo Salcedo
7^th February 2003

"Global optimization is a powerful tool that can be used with advantage over empirical methods to improve equipment design. I will show how this was achieved with cyclones and electrostatic precipitators (ESP's), two common dedusters in many industries, employed for the recovery and/or removal of particles from a gaseous stream. Cyclones, which are not very efficient but are not expensive, were designed for maximum efficiency while keeping costs (pressure drop) low. ESP's, which are efficient but very expensive, were designed for minimum annualized cost while keeping efficiency high. Two industrial applications of the optimized cyclones will be shown"

Authors: Ant�nio Ferreira and Nuno Faria
21^st March 2003

"Industry is more and more required to manufacture powders with defined properties interms of size and morphology. Among others, these characteristics affect filterability of slurry, flowability in storage tanks and the aspect of the final product.
The most important characteristic of the particle populations is undoubtly its size distribution. However, the characterisation of the particle morphology in systems like sucrose or salt crystals is an essential tool to understand the crystallisation mechanisms and the properties of the final product. Phenomena like agglomeration have a direct impact on the crystalisation kinetics, purity, crystal size distribution and their shape and final quality. In spite of the importance of this phenomenum to the crystallisation process, it had neve been quantified before.
In this presentation, a general description of some of the image analysis techniques used in the characterisation of sucrose and NaCl crystals will be made. Also, some of the results obtained shall be presented along with other
examples of possible applications of the image analysis technique."

Author: Jaime Villate
23^th April 2003

"The Large Hadron Collider (LHC) being constructed at CERN will be the world largest particle collider and it is expected to open new frontiers in particle physics. There will be 4 detectors along the LHC, dedicated to various different experiments. One of those detectors, ATLAS, will be used to search for some missing particles in the standard and supersymmetric models. The complexity of the ATLAS experiment poses several technological challenges that are being solved by a team of 1800 physicists and Engineers at several Institutions around the world. ATLAS depends strongly on the construction of software for detector simulation, data acquisition, storage and analysis.
A Portuguese team, involving the University of Lisbon, Catholic University at Figueira da Foz and FEUP, collaborates with the software groups of ATLAS. This talk will describe the contribution of the FEUP team, which is in the areas of databases design and GRID computing. 
The LHC, which is expected to start operating in 2007, will generate a volume of 5-8 PetaBytes of data per year. The storage and analysis of that huge amount of data exceeds the capacities of today's largest computers. To cope with that problem a global network (GRID) spread around research centers in Europe, Asia and America is being set up."

Author: Joana Peres
30^th June 2003

Artificial Neural Networks (ANNs) have been widely used for describing the reaction kinetics in biological systems embedded in hybrid models (e.g. Schubert et al. (1994), Montague and Morris (1994), Feyo de Azevedo et al. (1997)). Conventional Backpropagation (BP) and Radial Basis Functions (RBF) networks are the most employed architectures. One important issue related to the nature of the cell system is the fact that cells may process substrates through different metabolic pathways. This is the case of diauxic growth on two carbon sources. Or the case of aerobic/anaerobic growth depending on the presence or absence of dissolved oxygen in the medium. For example, S. cerevisae can grow trough three different metabolic pathways for exploiting energy and basic material sources and is able to switch between a respiratory metabolic state and a reductive metabolic state (Sonnleitner and Kappdli (1986)).
BPs and RBFs networks have some limitations for approximating discontinuous input-output systems. BPs tends to exhibit erratic behaviour around discontinuities (Haykin, 1994). RBFs are voted for local mappings but suffer from generalisation and intensive computation problems especially for resolution of fine details. There are strong reasons to believe that modular networks architectures may be advantageous for modelling reaction kinetics in biological systems. A modular network architecture consists of two or more (small) network modules mediated by a so-called gating network which decides how to combine their outputs to form the final output of the system. The learning of such networks is based on the principle of divide-and-conquer, i.e., the network modules compete to learn the training patterns. This type of architecture performs task decomposition in the sense that it learns to partition a task into two or more functionally independent tasks and allocates distinct networks to learn each task (Jacobs et al. (1991)). Microorganisms reaction kinetics are ruled by a rather complex network of metabolic reactions that can be viewed as being composed by a set of interconnected modules representing different pathways: glycolysis, TCA cycle, etc. Hence a modular network structure is hypothetically highly compatible with the internal structure of the system cell reaction kinetics.
The main objective of this study is to verify if modular network architectures, which are supposed to be able to perform task decomposition, are able to discriminate between reaction pathways in complex biological reaction schemes. The S. cerevisae yeast and the Waste Water Treatment serves as examples to illustrate the application of the Modular Networks.

Author: Vytautas Galvanauskas
21^st July 2003

Since modern biotechnological industry deals with very complex and expensive production processes, it is of advantage to optimize the processes in order to significantly reduce production costs, increase the safety of operation, maximize recovery and improve the quality of a product. From the point of view of biotechnological engineering, it is straightforward to improve the performance of a biotechnological process by improving process measurement, modeling, optimization and control. Each of the issues has its specific problems and bottlenecks. The issues are dependent on the development of such neighboring fields of science and industry as computer hardware and software, measurement and control devices and methods, and modern model based optimization and control techniques. The harmonious interplay between them, taking advantage of recent achievements in these areas and compilation of all (measurement, modeling, optimization, control) problems into one complete task, makes possible a significant improvement in biotechnological process performance. Furthermore, the development of a model based design procedure decreases the time consuming preparation and scale up steps of a biotechnological process. In this work the author has presented a procedure which combines the distinct issues of the model based optimization problem in fed-batch biotechnological processes. The mathematical models for the various typical fed-batch bioprocesses and of different complexity are built and analyzed. Successful structure definition of the mathematical models, and efficient incorporation of a priori knowledge into them led to good initial conditions for the optimization procedure and to more rapid convergence to an optimum. Investigation with simulation tests also showed that even relatively simple models can give to satisfactory estimation/ extrapolation quality. An investigation of the parameter influence on the process performance index was made. Comparison of different optimization techniques (Pontriagin's principle of maximum, modified Bellmann's dynamic approach, evolutionary programming technique) was performed. Evolutionary programming technique proved to be an effective and flexible tool for the optimization of fed-batch biotechnological processes. By means of the model based design procedures, involving only from the three to five process design iterations, it was possible to find quasi-optimal control profiles and initial conditions- to optimize the process.

Author: Cristina Oliveira
29^th September 2003

"The precipitation of calcium phosphate has been studied by various authors under different conditions. Depending on the precipitation conditions, like temperature, level of supersaturation, pH and initial concentration of reagents, one can obtain different calcium phosphate phases. One of them is the dicalcium phosphate dihydrate (CaHPO4.2H2O), hereafter named DCPD or brushite. This product, recognized as an important product in the application of fertilizers to soil, has been studied mainly for its role in the physiological formation of calcium phosphates. Previous works investigated, under different conditions and precipitation methods, what is the most stable calcium phosphate phase, what are the chemical species that precede the final one, the effect of impurities, and the kinetic laws governing its formation. Due to the very different conditions in which these previous works were made and taking into account the great complexity of this system, it is not simple and easy to estimate what happens in a very specific situation. The present study arose from doubts raised in a more vast investigation on dicalcium phosphate, concerning the precursors of brushite and how the intermediate phases developed into the final phase. In short, the present work studies the mechanisms that govern the formation of brushite under specific experimental conditions. The precipitation was carried out mixing equimolar quantities of calcium hydroxide and orthophosphoric acid. The experiments were performed for different initial reagent concentrations and always at 25 C. A model, divided into five stages, is proposed. This model explains all the precipitation process, identifying the existing phases at each moment, its evolution along all the process until the complete transformation into brushite. This kind of research is of particular interest for the chemical industry and for the understanding of calcium phosphate formation mechanisms. "

Author: Rui Oliveira
26^th April 2004

"The dynamical optimization of fermentation processes depends critically on the accuracy and predictive capacity of the underlying model. Mainly because of the inherent complexity of biochemical processes it may be very difficult and costly, if not impossible, to develop such a model based solely on first principles knowledge. The application of the hybrid modeling techniques may represent a cost-effective alternative for the analysis of such complex processes (Schubert et al. (1994), Preusting et al. (1996), Psichogios and Ungar (1992), Montague and Morris (1994), Feyo de Azevedo et al. (1997), Peres et al. (2001)). The most widely adopted hybrid structure combines material and/or energy balance equations with Artificial Neural Networks (ANNs). The ANNs are especially dedicated for the modeling of kinetics for cell growth nutrient uptake and products formation. Although the hybrid model synthesis requires less data than a single network for representing the complete process, a large amount of data is all the same required before the final model can be used for optimization. This is so essentially because of the problem of unreliable neural network estimates/predictions in extrapolation conditions. This represents a serious liability for this technique and restricts the target applications, i.e. only established processes with large amount of data available can be considered. In this work the hybrid mass balances-ANNs modeling is also adopted but a novel strategy is followed in the way the a posteriori dynamical optimization is performed, which overcomes the problem of unreliable network estimates. The optimization is realized numerically with a genetic algorithm but a penalty term is included in the objective function definition for unreliable network estimates. The main idea is to use a clustering technique to supervise the reliability of the network (Leonard et al. (1992)). The Gaussian clusters output can be interpreted as the probability of a given input vector belonging to the domain of experience of the network. The value of this probability is related to a given optimization risk. The value of the risk is used to trigger a penalty term in the objective function. Several simulation tests with a process with known optimal PMP (Pontryagins Minimum Principle) trajectories (Park and Ramirez (1988)), showed that this strategy can be applied in a batch-to-batch basis starting from a single process experiment. After a given optimization step a new feeding strategy is proposed for the next experiment and so on with progressive convergence to the optimal path. Thus the proposed method broadens the application of the hybrid technique to 'newer' processes with higher potential for optimization and with fewer data available than old processes. Currently this procedure is being used to optimize the acetate and ammonia feeding strategy in a pilot Polyhydroxyalkanoates (PHA) production process."

Author: Dennis Bonn�
19^th July 2004

"When in pursuit of optimal and reliable operation of batch processes, the limiting factor is often the lack of reliable models. This contribution presents a methodology for rapid acquirement of discrete-time state space model representations of batch processes based on their historical operation data. These state space models are parsimoniously parameterized as a set of local, interdependent AutoRegressive Moving Average models with eXogenous inputs. The modeling methodology is applied to sugar crystallization."

Author: Pedro Mena
13^th September 2004

"In bubble column reactors there are two principal flow regimes [1,2,3], the homogeneous (HoR) and the heterogeneous (HeR). These reactors have different behaviour in HoR and HeR, thus the dependences of the rates of mass, heat and momentum transfer on the design and operating parameters (such as reactor geometry, gas and liquid flow rates and properties of the contacting phases) are also very different. Therefore, for rational reactor design and operation it is of crucial importance to know the range of parameters over which a certain regime prevails and the regime transition conditions [4]. The goal of this work was to examine the influence of solid phase on homogenous regime stability and regime transition in a three-phase bubble column. For that, two studies were done, one focused on the flow regime transition and the other on visualization of bubble-particle interaction."