<
p>
For Phase 1, various feedstocks from coal, petroleum and bio-based acrylonitrile were investigated for their ability to produce carbon fiber within DOE strength (250 ksi), modulus (25 Msi), elongation (>
1%) and cost (?$5/lb) goals for wider utilization in light-weight vehicles for the automotive industry. The overarching approach was to chemically identify and map the various feedstocks, understand the molecular transformation to produce carbon fiber precursors (mesophase for pitch materials of polyacrylonitrile polymer) and how these impact carbon fiber quality. Predictive models were built to relate the feedstocks and precursors to carbon fiber properties using atomistic and micro-scale modeling within �15%. Using tow-level carbon fiber epoxy composites, carbon fibers tow-level predictions of elastic components of the carbon fibers were also successfully modeled within �15%, using finite element analysis to confirm single filament carbon fiber properties. These models were validated with tow-level epoxy composite experiments from carbon fibers produced in this study.<
/p>
<
p>
New approaches to machine learning were developed which can provide some chemical and physical insight behind correlations while also minimizing the risks of overfitting. This was accomplished by selectively including data from correlative heat maps, as opposed to using all collected data indiscriminately. The optimization of essential parameters feeding into the machine learning algorithms were tested and successfully validated against experimental data.<
/p>
<
p>
High temperature coal tar pitch, petroleum pitch and bio-based acrylonitrile were down selected for further study and scale up. The scaled up materials continued to meet DOE mechanical property targets. A more detailed cost analysis showed that all three feedstocks were capable of being produced at less than $5/lb within given assumptions of the various models.<
/p>
<
p>
In Phase 2 the tools and methods developed in the Phase 1 are integrated and further developed to relate the atomistic properties of CF feedstocks to single filament CF properties and ultimately to the tow-level composite fiber properties. This was accomplished through detailed understanding of the feedstock molecules, how these are changed to produce CF precursors (mesophase or polyacrylonitrile, PAN), and the further development of the chemical structure during CF processing. These aspects were then integrated using neural networks (NN) to develop the relationships to the various physical properties of the CFs. This was ultimately validated with limited data available on the scaled up samples of CF produced under this program. CF properties such as elastic modulus, compressive modulus, shear modulus and Poisson?s ratio could be predicted well within �15%.<
/p>
<
p>
For bio-PAN modeling efforts, the initial coarse-grain molecular dynamics simulations (CGMD) methods were found to give and overestimation of some elastic properties. This was corrected by modifying the CGMD models to add back in the hydrogen atoms and allowing all the atoms to relax, thus converting it back to a fully atomistic model. This method was also adapted so that various biomass-based polyacrylonitrile (bio-PAN) processing parameters such as stretching force, stabilization temperature and stabilization time could be used to predict CF properties, as well as define limits of CF failure during CF production. It is obvious that this type of novel modeling could be further developed to help guide PAN-based CF production to target various CF qualities without needing to take a trial and error approach. For tow-level CF properties the various elastic components and density were also easily predicted within �15%.<
/p>
<
p>
An important hurdle that was overcome for complete model integration was how to encrypt the CF processing data (spinning conditions and post processing of as spun fibers, e.g. stabilization and carbonization) without violating export control. This was important so that the processing, which is critical to CF properties, was not treated as a black box. A method was developed to encrypt the various CF processing parameters so they could be used for machine learning (ML) to develop web-based, user-friendly, platforms for predicting CF properties. Model integration showed that, based on the chemistry of the feedstocks and precursors, it is possible to predict elastic CF properties and tow-level composite properties well within �15%.<
/p>
<
p>
Model integration between the different scales (micro- to macro-level) was achieved using NN by considering the molecule properties and their connectivity during mesophase production and CF production (primarily through carbonization, although graphitization was considered), relating the localized molecular ?sheet? properties to the CF microstructure, and finally to tow-level epoxy composite CF properties.<
/p>
<
p>
Several Kg batches of petroleum pitch mesophase (PPM) were produced by Koppers. It was discovered that volatiles in some PPM batches were causing some inconsistencies in the spinning and CF properties. It was also demonstrated that primary carbonaceous insolubles, which was produced during the PP production, had a negative impact on CF properties. This was demonstrated by carefully filtering PP to remove the insolubles and producing PPM of the same softening point for the filtered and unfiltered samples. CF from these filtered and unfiltered (referring to the isotropic pitch, not filtration of the mesophase) PPM were produced using the same melt-spinning processes. PPM produced from filtered PP produced significantly smaller diameter CF with higher strength and elongation, but lower modulus. Additional characterization data for the PP and PPM samples were also collected during this work to be used in further developments of ML.<
/p>
<
p>
Overall, DOE targets for strength (250 ksi), modulus (25 Msi), elongation (>
1%) and cost at ?$5/lb were all met with scaled up bio-PAN and PPM CF during BP3. Several large strides were made during this program to address different facets of pitch-based and PAN-based feedstocks, CF production and integration of models for a holistic approach to arrive at tow-level predictive capabilities. However, it is admitted that given the large scope of the project various aspects are more developed and validated than others, and additional work is needed to fully understand all aspects of CF production.<
/p>