WP1 Project management
Establishment of a multidisciplinary research team with transversal skills. Organization of the selection of postdocs, enhancing the employment of young people with attention to equal opportunities.
WP2 – Modelling and Design of the device
Modelling and numerical solution of biphasic EHD problems. The tracing of the interface between the liquid and the gas is obtained using the volume of fluid method (VOF) without any simplification of the electrical behavior of the fluids involved. The physico-chemical properties of the system will be introduced as parameters, and parameter sweeps will be used to achieve control on the operating conditions to be selected for specific biological fluids
Selection of the model fluids for testing
Model biological fluids will be characterized to identify rheological properties to be inserted in the model to be simulated. Both healthy and pathological situations will be considered. As a benchmark, the viscosity of the body fluids will be measured at different temperatures with a cutting edge newly built desk micro-rheometer developed in our Labs [7].
Preliminary test with AI-in-the loop
Preliminary tests will be performed using model liquids of known parameters that have been selected and modelled to calibrate a general-purpose setup in which all geometric and functional parameters are reconfigurable. Definition of the relevant geometrical parameters of the evolving liquid drop at the base of jetting/dispensing and 3D deformation/elongation. Calibration setup will be driven by introducing “AI-in-the-loop” of PY modelling to select the optimal experimental parameters and to reduce time and cost of lab experiments with dataset developed in previous activities.
Development of an AI system based on synthetic training video flows
Synthetic video flows generated in previous activities will be used to design a novel AI model, i.e., a customized Convolutional NeuralNetwork (CNN), preliminary trained and tested only with synthetic samples for microfluid characterization/classification purpose. The pre-trained network will form the basis for processing the experimental data, achieved in WP3, that will act as a phase of fine tuning of the AI model.
Design of a compact PY device
The results of the previous tasks will be used to define the geometric and others parameters of the device: dimension and position of the crystal, range of temperature and heating source, sample slide geometry, position and surface properties, sample read-out and imaging set-up.
WP 3 – Fluids Testing and Data collection
This WP will be devoted to test the liquids and collect experimental data that will be used in the WP4 for the learning step. Based onthe investigation and design parameters of the WP2, the chosen fluids (types, volumes, time of evolution, cone formation, etc.) will be tested by using different manipulation modes: dispensing, elongation, and 3D shapes. The outcome of a preliminary model assessment and results of the simulations in the WP2 will allow to define the type of liquids to be tested, parameters of interest, and experimental systems settings.
Setting the experimental PY systems
Two different experimental set-ups will be arranged for actuating the proper manipulation that will work on liquid samples (smallvolume sessile drop). The first setup will be used for testing fluids for “high-speed” formation time (range 0-5s) and will focus on the dispensing or jetting modality. The second setup will be implemented for “longer” deformation times (range 5s-20s), allowing the 3D deformation of the sessile drop and eventually its elongation under the PY effect. The geometric aspect, size and thickness of the ferroelectric crystal will be defined according to the outcome of the simulation and theoretical model assessed in WP2. The PY activation will be designed and controlled to excite liquid manipulation on the different chosen liquids as per WP2. The results of 3.1 will be the input parameters for video data collection and will allow to select the setup for the pyro-rheological testing.
System for video capture in 2D and 3D (CNR)
It will be designed and realized an optical system to capture the displacement of the fluid under the action of the PY pressure. The system will exploit two different technologies for capturing experimental data of dynamic morphological changes of the liquids activated by PY pressure: i) 2D imaging system made of a LED light source, a CMOS camera and proper optics for imaging; ii) 3D-holographic imaging system to analyze complex 3D evolution. For both vision and capture systems that will be considered in Task 3.2; experimental parameters will be optimized in terms of resolution of the camera, frame rate of video acquisition, and modality for image capture and storage. A dataset consisting of videos will be built and passed to WP4
Investigating of different classes of liquids and data collection
Measurements of the rheological parameters of interest will be carried out with the PY device. Different biological liquids chosen inWP2 will be tested based on their rheological properties by the system designed and realized in the previous Tasks 3.1-3.2. A complex dataset will be built with the optics systems by considering different fluids and various PY pressures. Videos will be stored in appropriate format. The sequence of digital holograms will be processed by a proper suite of holographic reconstruction algorithms for tracking profiles of their 3D motion resolved in time and space.
WP4 – Machine learning of rheological parameters from image/video data
In WP4 different AI-models will be studied to automatically extract the relevant rheological information from the experimental data provided by the previous WPs.
State-of-the-Art literature analysis
In-depth literature review on the application of AI for microfluidics characterization/classification in order to assess the relevant works recently published on this cutting-edge subject of research.
Data preparation and preprocessing
The data of interest are derived in the different forms of images, video-flows, and vectors from the experimental parameters protocol. Both the format of the data and their proper size will be defined in this task, whose scope is to store the data to be processed by the neural schemes in a repository according to the specifications. The data will be pre-processed to reduce the impact of noise through signal processing.
Data encoding and decoding (Autoencoder)
Autoencoder-based models will be designed to generate a denoised representation of data and to extract some non-trivial features from combination of different parts of the images (videos). The defined latent representations can form the basis of a generative pipeline relevant to model classes of data potentially significant but absent or misrepresented in the databases at hand.
Convolutional neural networks (CNN) for image processing
The customed CNN proposed in Task 2.4 will be tested and validated on experimental image data. In addition, suitable CNN schemes will be also used to process 3D data from movie, thus incorporating dynamic information on the flow evolution under the PY excitation.
Classification of different fluids considered
In order to improve the classification of different fluids, the CNN scheme will be enriched by using movie data and introducing additional information coming a priori or from engineered features to the serialized feature maps.
Dynamic analysis and representation of fluids’ behavior through holograms
Self-supervised learning methods based on deep CNNs and specific loss functions will be also developed to understand the dynamics of the fluid time-evolution and capture hidden aspects of the holograms.