1、Introducing ekional source model to account for some of the problems mentioned.
2、Rupture which is initially allowed at any point on the infinite fault plane is firsr restricted to this zone. Secondly, we allow rupture only happen within a certain radius R away from the previously defined point source location labeled center point. The radius is the only variabe in our model controlling the size of the rupture surface.
Kiwi has mentioned the disadvantages of traditional dscretization includes these errors:
1、 Each point in the sub-fault incorrectly radiates the same Green's function waveforms; (它会影响最低频率的滤波)
2、The temporal shifts of the signal radiated by different parts of the subfaults are neglected;（难道不是体现为破裂传播速度的不同吗？）
3、Suppression of local inhomogeneities of the moment tensor density inside the sub-fault.
Give a simplerule about the value of rupture velocity; And the size of sub-faullt is made dependent on the maximum frequency and the maximum slowness.
这里的Greenfunction内插方法使用的是bilinear interpolation, module is 'interpolation'; The API to read and write green's funtion database and the Green's function catche is implemented in 'gfdb';
Conduction of filtering and tapering:
这些是用以抽取特定的震相到时部分，并将其限定到特定的频率区间；
Tapering 是先于 filtering的，以防附近更强的震相由于滤波的影响不小心进入分析窗口。
反演方案：
由于一共有13个待解参数，想要对这些参数全部执行格网搜索是不太可能的，因此我们将搜索空间分成很多个子空间（子空间之间可以部分重叠），然后在这些子空间中用格网和梯度法的结合进行搜索。
为什么可以使用子空间（即分开反演各个参数）：
Because the misfit function can be designed in a way, which makes it independent of some parameters. For example, if we base the misfit calculation on amplitude spectral, it becomes insensitive to the exact location and time of the centroids, or by low-pass filtering the seismograms it becoes less dependent of the parameters dealing with the fault extension.
从而，反演步骤可以设定如下：
首先，点源参数由地震波的低频成分所确定。接着，点源参数被固定，余下的参数在更高的频率被反演。
台站质量评价的策略：
首先，获取一定时间段内，震级在某个范围的所有地震事件的地震数据。
接着，这些数据被预处理：
接着，这些地震事件所对应的理论地震波被计算。（By forward modeling synthetic seismograms for a large ensembles of sources representing a given a prior expectation on the source parameters, a simplified but practical application of this is to forward modeling point sources of various strike, dip, rake combinations.接着，将观测波形设定为0，同时使用与反演中同样的Ttapering 和 filtering, 然后 计算the trace misfit in L1-norm way,并取平均值，做倒数，作为该台站的权重 ）；这里用的filtering 是 0.005-0.02HZ;
接着，首先将其他点源参数由Global CMT所公布的参数确定后，Centroid time 和 rise-time 通过格网搜索法 被调整。然后根据调整后的time,计算正演台站波形想要和观测波形最佳拟合的移动时间，并由此计算所有事件的平均misfit ，并可以根据每个台站计算出来的值画一个misfit curve, 当某个台站的记录时间出错时，被影响的那些台站会有一个小的时移。而当一个台站信号很差时，misfit curve的形状会被扰乱。这些操作执行完后，基于这些误差曲线的“stream badness”被用来作为一个判断标准，即当在misfit最小时的时间移动小于一个阈值时，这个misfit被称为stream badness。