Page 17 - IJAMD-1-1

International Journal of AI
for Material and Design ML in 3D bioprinting of cultivated meat

summed to give the final output. The SISO model achieved analysis demonstrated potential as an efficient and
mean errors of 3.49 – 10.15%, while the MISO model non-destructive technique for predicting the sensory
achieved mean errors of 2.57 – 9.44%, demonstrating the quality traits of beef.
ability of ANNs to account for complex flavor interactions Creating useable PLSR models also depends on the
to achieve a higher level of accuracy. Aside from this study, analytical technique used, with some analytical techniques
limited research has been conducted on the application of giving significantly better correlations even when all else
ANNs for the prediction of meaty flavors in solid foods, is being held equal. Another study applied PLSR to time-
with most other ANNs being applied to flavors in beverages domain nuclear magnetic resonance (TD-NMR) to predict
instead.
three sensory attributes of grilled beef, using signals from
3.2. Prediction of meat flavor from sensory array data Carr-Purcell Meiboom-Gill (CPMG) and continuous
wave-free precession (CWFP) decay. The correlation
The other branch of ML in meat sensory analysis primarily coefficients predicted by PLSR models gave magnitudes
relies on non-chromatographic techniques, predominantly of approximately 0.71, 0.56, and 0.67 for flavor, juiciness,
spectroscopy, electrochemical sensing, or a combination and tenderness, respectively, with the CPMG sequence,
of both. Commonly, principal component analysis (PCA) and 0.8 for cooking loss, 0.99 for fat content, and 0.73 for
and PLSR are employed to unravel the complexities of these moisture content with the CWFP sequence. The results
63
multivariate data. However, due to the non-linear nature indicated the feasibility of concurrently assessing multiple
of the relationship between sensory attributes with data sensory properties by utilizing TD-NMR signals from
from gas sensor arrays and electrochemical sensor arrays, two distinct sequences, specifically CPMG and CWFP,
SVM and backpropagation neural networks (BPNN) have through a PLSR multivariate model. The validation
gained traction in recent applications (Table 2).
process, which demonstrated low root mean square errors
The ability of PLSR to predict meat flavor from (RMSE) for both the calibration and validation datasets,
spectroscopic or electrochemical data, while potentially provides substantial evidence for the ability to predict the
accurate, is also very dependent on the specific cut of considered parameters with precision. This finding implies
meat and the data processing applied. In one study the potential reliability and accuracy of the models in
utilizing PLSR on Raman spectroscopic data to predict forecasting sensory attributes based on TD-NMR signals
sensory attributes of cooked aged silverside beef, a large obtained from these sequences.
difference was observed between its predictive ability Various spectroscopy techniques, including visual
for flavor-related attributes compared to texture-related reflectance (VIS), near-infrared (NIR) reflectance, Raman
attributes. The attributes related to aroma and taste had scattering, fluorescence emission, and low-field H nuclear
1
low R values ranging from 0.19 to 0.34, whereas the R magnetic resonance (NMR), were individually employed
2
2
values for texture and mouthfeel-related attributes were alongside PLSR to predict the sensory attributes of cooked
notably higher at 0.62 to 0.71. This finding suggests that pork. The PLSR analysis of VIS data revealed the highest
61
Raman spectroscopy holds promise in predicting the correlation coefficients for metallic, sweet, sour, and
sensory-perceived quality of beef, particularly as texture
and tenderness emerge as dominant factors influencing linseed oil-like attributes, ranging from r = 0.92 to 0.93.
For NIR data, the highest correlation coefficients were
the overall acceptability of beef. In another study, PLSR r = 0.93 and 0.94 for sour and monosodium. In the case of
was used on Raman spectroscopic data to predict sensory fluorescence data, the strongest correlations were found for
attributes of a different cut of grilled aged beef. It achieved rancid and linseed oil-like, with coefficients of r = 0.79 and
a more similar range of cross-validated coefficients of 0.80. Regarding NMR inversion recovery data, significant
determination (R 2 CV ) between flavor-related and texture- correlations were noted for sour and monosodium, with
related attributes, with R 2 CV of 0.50 to 0.76 for texture and
mouthfeel-related attributes and R 2 CV of 0.52 to 0.84 for coefficients of r = 0.91 and 0.94, respectively. NMR (CPMG)
aroma and taste-related attributes. When PLSR models exhibited high correlation coefficients of r = 0.89 and 0.90
64
were subsequently constructed on subsets of beef samples, for salt and monosodium. Hence, with PLSR, depending
such as those with identical age or breed, significant on the sensory attribute to be predicted, the meat species,
enhancements in predictive capabilities were observed for and the cut of meat, the most optimal and appropriate
overall sensory attributes. Specifically, R 2 CV values ranged analytical technique will need to be empirically decided,
from 0.63 to 0.89 for each breed type and from 0.52 to likely needing new datasets to be generated whenever any
0.89 for each age group. Chemometric analysis exhibited variables or procedures change.
62
robust correlations among sensory attributes, and the Aside from PLSR, SVMs have also been employed to
integration of Raman spectroscopy with chemometric predict beef flavor grading using data from 12 ion-selective

Volume 1 Issue 1 (2024) 11 https://doi.org/10.36922/ijamd.2279