Hydrothermal polymerization of porous aromatic polyimide networks and machine learning-assisted computational morphology evolution interpretation

We report on the hydrothermal polymerization (HTP) of polyimide (PI) networks using the medium H2O and the comonomers 1,3,5-tris(4-aminophenyl)benzene (TAPB) and pyromellitic acid (PMA). Full condensation is obtained at minimal reaction times of only 2 h at 200 °C. The PI networks are obtained as monoliths and feature thermal stabilities of >500 °C, and in several cases even up to 595 °C. The monoliths are built up by networks of densely packed, near-monodisperse spherical particles and annealed microfibers, and show three types of porosity: (i) intrinsic inter-segment ultramicroporosity (<0.8 nm) of the PI networks composing the particles (∼3–5 μm), (ii) interstitial voids between the particles (0.1–2 μm), and (iii) monolith cell porosity (∽10–100 μm), as studied via low pressure gas physisorption and Hg intrusion porosimetry analyses. This unique hierarchical porosity generates an outstandingly high specific pore volume of 7250 mm3 g−1. A large-scale micromorphological study screening the reaction parameters time, temperature, and the absence/presence of the additive acetic acid was performed. Through expert interpretation of hundreds of scanning electron microscopy (SEM) images of the products of these experiments, we devise a hypothesis for morphology formation and evolution: a monomer salt is initially formed and subsequently transformed to overall eight different fiber, pearl chain, and spherical morphologies, composed of PI and, at long reaction times (>48 h), also PI/SiO2 hybrids that form through reaction with the reaction vessel. Moreover, we have developed a computational image analysis pipeline that deciphers the complex morphologies of these SEM images automatically and also allows for formulating a hypothesis of morphology development in HTP that is in good agreement with the manual morphology analysis. Finally, we upscaled the HTP of PI(TAPB–PMA) and processed the resulting powder into dense cylindrical specimen by green solvent-free warm-pressing, showing that one can follow the full route from the synthesis of these PI networks to a final material without employing harmful solvents.


Chemicals and equipment
All chemicals were commercially available and used as received from TCI without any further purification. Deionized water was used at all times. As reaction vessels Berghof Digestec system and Parr steel reactors were used. 1 H NMR spectra were recorded on a Bruker AC600 spectrometer. Chemical shifts are reported in ppm (δ) relative to tetramethylsilane and calibrated using solvent residual peaks. Data are shown as follows: Chemical shift, multiplicity (s = singlet, d = doublet, t = triplet, q = quartet, quin = quintet, m = multiplet, b = broad signal), coupling constant (J, Hz) and integration.

Monomer salt synthesis
TAPB (1 eq., 1 mmol) and PMA (1.5 eq., 1.5 mmol) were weighed in to a round bottom flask and suspended in 10 mL of deionized water. The white suspension was stirred at room temperature overnight. In the course of the reaction, the suspension turned green. The solid was filtered, washed with water and very little acetone and subsequently dried in vacuo. Yield: 785,3 mg (54%). 1

Hydrothermal polymerization
TAPB (1 eq., 0.2 mmol) and PMA (1.5 eq., 0.3 mmol) were weighed into a glass liner and suspended in 10 mL H2O. After ultrasonication for 10 min, the glass liner was sealed in a steel autoclave and heated to 200 °C or 250 °C for 2-168 h under autogenous pressure (~18 bar). The brownish precipitate was filtered via a Büchner funnel (13-15 μm particle retention) and washed with water, acetone (techn.) and THF. Afterwards 75 mg of the product (ranging from brown powder to spongy monolith depending on the conditions) was dried at 80 °C in a vacuum oven overnight.

Microwave synthesis
Monowave TAPB (1 eq., 0.2 mmol) and PMA (1.5 eq., 0.3 mmol) were weighed into a microwave reactor (Anton Paar, G30) and suspended in 10 mL H2O. After adding a stirring bar, the reactor was closed with a cap including a PTFE septum and placed in the Anton Paar Monowave 400. The vessel was heated as fast as possible to 200 °C, stirred with 600 rpm, and held at that temperature for 4 h. Upon heating the reaction mixture turned first green, then blue and finally, when the temperature was reached, a brown precipitate was formed. Afterwards, the reactor was cooled down to 70 °C and the brown precipitate was collected via filtration. The product was washed with water, acetone, ethanol and THF. Finally, the product (brown powder, 73 mg) was dried overnight, in vacuo, at 80 °C.

Multiwave Pro
TAPB (1 eq., 0.8 mmol) and PMA (1.5 eq., 1.2 mmol) were weighed into a microwave reactor (Anton Paar, N8QX), suspended in 40 mL H2O (degassed with Ar) and a stirring bar was added. In total four of these reactors were put in to the Anton Paar Multiwave Pro and heated within 20 min to 250 °C and held there for 4 h while stirring on the highest level. Afterwards, the reactors were cooled down to 70°C using compressed air. The reactors were opened, the brown powders of all four reactors were collected via filtration, washed with water, EtOH and acetone. The procedure was repeated four times and the 16 batches were combined. The product was then dried in vacuo. Yield: 7.364 g.

Solid state polymerization
500 mg of [H3TAPB 3+ PMA 2-] were placed in a round bottom flask, containing a stirring bar and evacuated. Afterwards, while continuously applying vacuum, the flask was placed in an oil bath which was set at 200 C° for 8 h while stirring vigorously. A color change from pale green to yellow and finally to beige was observed after several hours of heating. When cooled down to rt, the beige solid powder (360 mg) was collected and stored.

ATR-FT-IR
IR-measurements were recorded with a PerkinElmer UATR Two FT-IR spectrometer and the data was processed with the PerkinElmer Spectrum software. Figure S 2 ATR-FT-IR spectra of the monomer salt (green, aryl-ammonium modes at 2876 and 2592 cm -1 ), a PI sample (tr = 24 h, Tr = 200°C; symmetric and asymmetric C=O stretches at 1775 and 1720 cm -1 as well as the C-N stretch at 1365 cm -1 ) and the two precursors TAPB (free amine modes at 3400-3300 cm -1 ) and PMA (C=O stretches).

Thermogravimetric Analysis
Thermogravimetric analysis was carried out using a Netzsch TG 209 analyzer at a heating rate of 10 K min -1 under nitrogen atmosphere, equipped with NETZSCH Proteus (Version 4.3) software.

Scanning Electron Microscopy
Scanning electron microscopy was carried out with a Quanta 200F FEI microscope. Typically, the samples were measured at 5 kV, with a working distance of 9 mm and spot size 2.0. Prior to imaging, samples were loaded onto carbon-tape on steel sample holders and coated by sputtering with a 17 nm thick layer of Au/Pd 60/40 alloy with a Quarum Q105T S sample preparation system. The 437 SEM images used for both the expert interpretation and the machine-learning supported morphology analysis are provided as an open access dataset with doi 10.5281/zenodo.4544904.       Low pressure CO2 physisorption isotherms were measured volumetrically at 195 K and 273 K up to 1 bar using an Autosorb-IQ-MP from Quantachrome equipped with a Quantachrome CryoCooler for temperature regulation. Isotherm points chosen to calculate the BET surface area were subject to the consistency criteria detailed by Rouquerol. 1 The pore size distribution was derived from the adsorption isotherms at

Hg-Porosimetry
The pore size distribution of sample ML129B (PI(TAPB-PMA), tr = 24 h and Tr = 200 °C) was determined by mercury intrusion porosimetry (PASCAL 140/440, Thermo Fisher Scientifc), assuming a contact angle of 140°. The maximum intrusion pressure was 400 MPa, corresponding to a minimum accessible pore opening diameter of 4 nm.
Density and pore characteristics of the warm-pressed PI(TAPB-PMA) pellet were characterized by mercury intrusion porosimetry [same testing conditions as other sample]. Owing to an apparent compression of the sample at intrusion pressures above 100 MPa, a compressibility correction was applied to the intrusion curve. Furthermore, the flexural strength of the warm-pressed PI(TAPB-PMA) material was evaluated using a three-point flexural test setup, following EN ISO 178. Test specimens with dimensions of 5 x 1.6 x 30 mm were obtained using a diamond cutting disc. A total of three specimens were tested until fracture using a crosshead speed of 1 mm min -1 (Universal testing machine Model 1474, Zwick, Germany).

Powder X-Ray Diffractometry
A PANalyticalX'Pert Pro multi-purpose diffractometer (MPD) in Bragg Brentano geometry operating with a Cu anode at 45 kV, 40 mA and an X-Celerator multichannel detector was used. Samples were in most cases ground and mounted as loose powders on silicon single crystal sample holders, in other cases the spongy structure had to be fixed to the sample holder with heptane/Vaseline solution. The diffraction patterns were recorded between 1 and 30° (2θ) with 74.970 s/step and a step size of 0.0201°, sample holders where rotated during the measurement with 4 s/turn.

Differential Scanning Calorimetry DSC
DSC measurements were performed on a DSC832 Mettler Toledo machine with a heating rate of 5 K/min and a nitrogen gas flow of 10 µl/min. The temperature ranged measured was 30-400 °C. PI-pellet Up-scale PI 3) as an interface for artificial neural networks. In a preliminary step images were split in an ordinary grid into 20 tiles. Each individual tile was then analysed whether it contains useful information (foreground). This was performed by first calculating a pixel foreground threshold using the Otsu method and then removing tiles with less than 10% foreground pixels. Next, we used the tiled images as input for six different artificial neural networks: Xception, VGG16, VGG19, InceptionV3, MobileNet and ResNet. For all six neural networks the prediction layer has been removed, such that the final output of each network is a pseudo feature string varying between 512 and 2048 dimensions. All six neural networks have been pre-trained on regular images using the imagenet dataset. The performance of individual network architectures as well as different feature normalization methods were evaluated by calculating the Cohen's d (effect size) between the eucledian distances of tiles that belong to the same original image, and tiles that belong to different images.

Literature Comparison
Next, we calculated the centroid (mean) point within the 20 tiles of an image, such that each SEM image corresponds to exactly one point. To further analyse and visualize our data, we applied dimensionality reduction techniques to our dataset to reduce the number of dimensions to two. Both principal component analysis (PCA) and tstochastic neighborhood embedding (t-SNE) have been performed using the python package scikit-learn (v. 0.2.4).
To use a homogenous set of images with a similar range of magnification we decided to include only images between 1300x and 4000x which correspond to ca. half of all images. Next, we identified the ideal number of clusters using the k-Means clustering and the silhouette method. K-Means clustering partitions n datapoints, e.g., images, into k clusters in which each point belongs to the cluster with the nearest mean. The silhouette method calculates how similar objects are to their own cluster compared to other clusters. Values close to one indicate perfect cohesion in individual clusters, whereas values close to zero indicate bad clustering.