Application of Machine Learning Methods to the Open-Loop Control of a Freeform Fabrication System

Freeform fabrication of complete functional devices requires the fabrication system to achieve well-controlled deposition of many materials with widely varying material properties. In a research setting, material preparation processes are not highly refined, causing batch property variation, and cost and time may prohibit accurate quantification of the relevant material properties, such as viscosity, elasticity, etc. for each batch. Closed-loop control based on the deposited material road is problematic due to the difficulty in non-contact measurement of the road geometry, so a labor-intensive calibration and open-loop control method is typically used. In the present work, k-Nearest Neighbor and Support Vector Machine (SVM) machine learning algorithms are applied to the problem of generating open-loop control parameters which produce desired deposited material road geometry from a description of a given material and tool configuration comprising a set of qualitative and quantitative attributes. Training data for the algorithms is generated in the course of ordinary use of the SFF system as the results of manual calibration of control parameters. Given the large instance space and the small training data set compiled thus far, the performance is quite promising, although still insufficient to allow complete automation of the calibration process. The SVM-based approach produces tolerable results when tested with materials not in the training data set. When control parameters produced by the learning algorithms are used as a starting point for manual calibration, significant operator time savings and material waste reduction may be achieved. Introduction Solid freeform fabrication (SFF) is the name given to a family of manufacturing processes which allow three dimensional printing of arbitrarily shaped structures, directly from computer-aided design (CAD) data. (a) (b) Figure 1. Fabrication platform: (a) 3-Axis gantry robot for deposition with cartridge/syringe tool, (b) continuous wire-feed tool Typically, an SFF system consists of a tool which dispenses a “road” or line of material, a robotic positioning system which moves the tool along 3-dimensional trajectory, and a software control system. SFF has traditionally focused on producing passive mechanical parts. Advances in this technology and developments in materials science make it feasible to begin the development of a single, compact, robotic SFF system – including a small set of materials – which can produce complete, active, functional electromechanical devices, e.g. mobile robots. A research SFF system with two material dispensing tools has been constructed (Figure 1Error! Reference source not found.) pursuant to this goal, and Figure 2 depicts a Zn-air battery produced with this system. One of the challenges in developing such a system is in achieving precise, accurate, and repeatable dispensing of materials despite the difficulty of automatically monitoring output quality, and the significant variations in properties between materials – even between batches of the same material. Currently, these challenges are handled via an extensive and laborious manual calibration process for each batch of each material, immediately prior to use. During calibration, the SFF system produces a series of rectilinear test patterns (Figure 7), and control parameters are tuned until the produced pattern matches the desired pattern to the operator’s satisfaction. Figure 2. Zn-air battery produced via SFF The control parameters (Table I) describe piecewise-linear profiles (Figure 3) for the commanded extrusion rate from the tool and for the robot trajectory speed. The k-Nearest Neighbor and SVM algorithms have been applied to the computation of control parameters for the syringe tool (Figure 1Error! Reference source not found.a). The inputs (i.e. “attributes”) to the learning algorithms are simple quantitative and qualitative descriptions of the material, the tool, and the deposited road (Table II). The attributes consist of parameters which are strictly dependent on the material itself as well as desired properties of the extruded material road (e.g. width and height of the dispensed material). In that these attributes have been intuitively selected, it is unclear whether they are necessary or sufficient to fully represent the problem domain. Table I. Example control parameters vector Parameter Value Unit ACCELDELAY 0.2 s INPUTACCEL 1 step/ms 2 PUSHOUTSPEED 300 step/s PUSHOUTTIME 0.4 s INPUTSPEED 50 step/s OUTPUTACCEL 40 mm/s 2 OUTPUTDECEL -30 mm/s 2 OUTPUTSPEED 5 mm/s DECELDELAY 0.25 s INPUTDECEL 1 step/ms 2 PULLBACKSPEED -300 step/s PULLBACKTIME 0.4 s Tool and Robot Speed Profiles -4 -3 -2 -1 0 1 2 3 4 5 6