Page 17 - IJB-3-2

Samuel Charles Sklare, et. al.

3.2. Integrated system data
Continuously collected sensor data is put into a da-
ta base of “printing metadata” and subsequently ana-
lyzed. The goal of this data collection is to build a
se ries of machine-learning and data analytics tools
to further automate printing and help us learn from
every successful and unsuccessful transfer attempt. A
successful transfer occurs when a targeted subsection
of cell-laden material is transferred to the receiving
substrate, without blowing apart cells, and ejecting rather
than delaminating cells.
Using different materials for LDW will alter the op-
timal printing parameters; thus, every combination
Figure 5. Interactive grid-printing prompt
requires careful optimization to narrow the operating
parameters’ space. This information was formerly stored
in protocols and entered into printing software for each
specific session. Now this information is stored in a
database, the Pandas Python module is used to analyze
it and appropriate parameters for different cell types and
materials are automatically loaded. The stored collected
data from a printing session is now tagged with cell
types and material composition, in addition to the raw
sensor data (laser energy, distance between ribbon and
substrate, aperture opening, tempe rature, humidity, and
before/after pulse pictures). When printing sessions
include “metadata” collection, it is necessary for the
users to judge and indicate if each transfer is successful.
Experienced users can tell from the characteristic
abla tion bubbles on the print ribbon if the transfer is
successful for the specific LDW system in use. The most
reliable method is to move the ribbon out of the way and
focus the camera on the substrate. The examination of
Figure 6. Automatically-generated graphical guide to grid
program the substrate is not yet automated, and this dramatically
slows down the printing process. Therefore, printing
sessions specifically to collect this data are sometimes
low volumes of transferred biomaterials (usually cell- desirable. Combining this measured outcome (binary
laden biomaterials); essentially, LDW can be used as data) with the metadata creates a clear picture of de-
a rapid prototyping system for tissue constructs and sirable parameters and allows quantitative learning from
spatially defined biological experiments. every printing trial.
The current iteration requires several different para- Automating the step of classifying prints as successful
meter files to be edited and loaded in a specific order. or unsuccessful is being implemented using machine
It leans toward monitoring the system (environmental learning. The database of manually classified prints is
and operational status) and offering powerful semi- being built largely to facilitate this process. Combining,
automation rather than full-stack automation. Future binning, and manually classifying pictures of the ribbon
iterations of the control system should include remote and substrate before and after printing allows the
design and operation capabilities. A 3D model of a construction of an automated classifier using automated
micro machined feature and grid-printing routine could feature extraction and a neural network.
be created on a remote computer and then transmitted Integrated data collection and machine learning will
over the internet to the LDW system. Then, a technician soon be used to study the entire printing process and
could prepare the appropriate print ribbons, load the downstream experiments. An example implementation
ribbons into the machine, and start the automated pro- involves a simple live/dead experiment:
cess. Such a decentralized design approach could allow 1. This is a single-cell precision experiment; a sparse
many more researchers to use the same machine and ribbon is prepared. After preparing the print ribbon
increase the speed of the prototyping process. and substrate, the ribbon is scanned and automatically

International Journal of Bioprinting (2017)–Volume 3, Issue 2 105